A transparent Matrix-inspired figure presenting a red pill and a blue pill over live digital rain
 
Move toward a pill · the closer you get, the brighter it becomes
Agent Smith // System Audit
System anomaly detected
Mr. Anderson…
Evidence channel selected
SF BAY AREA · STAFF PM · TIKTOK (BYTEDANCE)
> STAFF PRODUCT MANAGER · AI/ML PRODUCT BUILDER

0→1→NShipped andMeasured.

I’m Tanishq Mishra, a Staff PM at TikTok with 8+ years across engineering and product leadership.

I work on products where the hard part is not shipping a feature, but deciding what the system should do, what it should never do, and which signals are strong enough to trust at scale.

The distinction

I don’t default to AI. I use it when the product problem genuinely requires probabilistic reasoning, unstructured data, or adaptive decisioning. I build products where AI is the core capability.

Problem framing · architecture choice · evaluation · guardrails · human fallback · latency · cost · measurable value
Get in touch LinkedIn GitHub
CERTIFIED PRODUCT LEADERSHIP
Product School PMC Certified badge
Product School PMC
Scrum Alliance Certified Scrum Product Owner badge
Scrum Alliance CSPO
MarkStrat Certified badge
MarkStrat Certified
Tanishq Mishra in a black leather jacket
📍 San Francisco Bay Area
Product depth
01PRODUCT

AI / ML Systems

AgentsRAGRecommendationsComputer VisionAnomaly DetectionEvaluation Frameworks
02PRODUCT

Payments Infrastructure

Credit CardsStablecoinsBNPLAccount UpdaterACHRTP
03PRODUCT

Risk & Decisioning

Risk Scoring ModelsFraud DecisioningIncident ManagementStep-Up VerificationAnomaly DetectionHuman Review
04PRODUCT

Developer & Data Platforms

APIsSDKsWebhooksCDPTokenizationIntegrationsAuthenticationPlatform Services
Flagship Product Projects

A live AI product, built end to end

A live portfolio assistant combining product discovery, system design, evaluation, guardrails, and hands-on implementation.

Product Project · Live on this site

The RAG chatbot in the corner? I built and shipped it.

It's a product build over my own work history: discovery, retrieval design, evaluation criteria, launch guardrails, and iteration. A Cloudflare Worker handles authentication and rate limiting; the retrieval layer grounds every answer before the LLM responds.

Visitor question→ Cloudflare Worker→ Retrieval index→ LLM→ Grounded answer
Problem framingRAG architectureEval set GuardrailsLive iteration
Fun facts

Built by Tanishq, powered by Groq through a Cloudflare Worker, with no model key in the browser and a dynamic 500-token ceiling for richer recruiter conversations.

Guardrails

Evidence-bound personal claims, metric integrity, prompt-injection refusal, confidentiality boundaries, and a direct human handoff.

Benchmark targets

60-question golden set · ≥95% fact accuracy · 100% link and metric accuracy · ≥95% injection resistance · p95 <5s.

How I think about products

From ambiguity to impact

A working AI product lifecycle, not a framework poster. Each stage answers a different decision question, and the detail changes with the product, not the discipline.

  • User research
  • Historical data
  • Behavioral data
  • Support signals
  • Market context
  • Existing system performance
  • Problem definition
  • JTBD
  • User pain
  • Business objective
  • Constraints
  • Success criteria
  • Baseline
  • Solution options
  • AI vs rules vs workflow
  • Data strategy
  • Human-in-the-loop
  • System boundaries
  • Failure modes
  • Guardrails
ValueBusiness impact · user impact · strategic alignment · time to value
FeasibilityImplementation complexity · technical feasibility · data readiness · integration dependencies
Economics & riskCost to serve · operational burden · model risk · latency · reliability · reversibility
LeverageScalability · capability building · platform reuse · future use cases
  • Core hypothesis
  • Smallest viable scope
  • Critical workflow
  • Minimum data requirement
  • Baseline model / rule
  • Acceptance threshold
  • Human fallback
  • Instrumentation
  • Validation plan
  • PRD
  • Architecture alignment
  • Data pipelines
  • APIs
  • Model integration
  • Evaluation framework
  • Guardrails
  • Experimentation
  • Rollout plan
  • Outcome metric
  • Adoption
  • Quality
  • Reliability
  • Unit economics
  • Operational burden
  • Customer value
  • Error analysis
  • Feedback loops
  • Retraining / iteration
  • Automation opportunities
  • Rollout expansion
  • Monitoring
  • Cost optimization
  • Capability reuse
1 / 7

Case studies, written as experiment readouts

Because that's how the work actually happened: a North Star metric, its drivers, and the counter-metrics that kept us honest.

Product walkthroughAgentic support experience for users and creators · 62% self-serve resolution
01 · PROBLEM

Scale support without trading away answer quality.

The existing support model depended heavily on human capacity. The product needed to resolve repeatable questions quickly while keeping policy sensitive and ambiguous issues out of unsafe automation.

02 · PRODUCT DECISION

Ground answers first, then add agency.

I chose retrieval augmented generation over free form generation so answers could be grounded in approved policy and help content. Intent routing decided when the system could answer, when it needed tools, and when a person should take over.

03 · BUILD

Design the workflow around confidence and escalation.

The product combined retrieval, intent classification, conversation context, source grounding, human escalation, and a feedback loop from agent corrections. The goal was a reliable support workflow, not simply a chatbot interface.

04 · EVALUATION

Measure resolution and failure quality together.

Offline evaluation covered intent precision and recall, answer grounding, escalation behavior, and common failure modes. Rollout was staged so we could compare self service gains against false escalations and customer satisfaction.

05 · READOUT

The product improved both automation and service quality.

62%Self service resolution
89% / 84%Intent precision / recall
28% → 11%False positive escalations
+8%CSAT
$300KAnnualized support cost savings
06 · LEARNING

The right automation boundary matters as much as the model.

The strongest product decision was not to maximize automation. It was to automate high confidence work, preserve human judgment for uncertainty, and continuously use those escalations to improve the system.

07 · TRADEOFFS & RISK

I accepted a lower automation ceiling to protect trust.

The tradeoff was speed of deflection versus answer safety. A more aggressive automation threshold could have resolved more tickets immediately, but it would also increase the chance of confident, policy-sensitive mistakes. I accepted more human escalations early because those cases protected the user experience and created labeled feedback that could safely expand automation over time.

Risk acceptedHigher short-term support load in exchange for safer learning and a more defensible path to scale.
Key decisions & metrics

Scale resolution only when answer quality, user trust, service quality, and unit economics stay healthy.

Key decisionOptimize the automation boundary, not maximum containment

Use retrieval and confidence based routing for repeatable questions. Escalate ambiguous, policy sensitive, or low confidence cases instead of forcing an answer.

North StarSelf service resolution with verified quality

62% self service resolution is the headline outcome. It only counts as healthy when intent quality, grounding, escalation behavior, and customer satisfaction remain within guardrails.

62% self service89% precision84% recall
GuardrailsTrust, service quality, and economics

Observed guardrails include false positive escalations and CSAT. Also track grounded answer rate, unsupported answer rate, p95 latency, escalation SLA, and cost per resolved conversation.

28% → 11% false escalations+8% CSAT$300K annualized savings
Scale gateExpand by intent cohort

Scale only after a cohort clears quality and escalation thresholds. If resolution rises while grounding, CSAT, or safety worsens, hold or narrow the automation boundary.

01 · PROBLEM

Creator cash flow was the product problem.

LIVE earnings can arrive in irregular bursts while creators still have recurring business expenses. The product opportunity was to shorten the gap between earning and usable funds, while giving creators a cleaner separation between business and personal money.

02 · PRODUCT DECISION

Treat access to earnings as part of creator monetization.

The UK product pairs TikTok LIVE rewards with a Visa debit card and business account. That turns payout access from a back-office settlement detail into a visible creator experience.

03 · BUILD

Connect rewards, card access, and spendability.

The Creator Card lets eligible LIVE creators access earned funds and spend them without waiting for a traditional bank payout cycle. The broader payments platform also depends on wallets, authentication, tokenization, settlement, and market-specific rails.

04 · ROLLOUT

Sequence by regulatory and operational readiness.

Payments cannot scale through one identical market implementation. Launches need the right partner setup, regulatory readiness, authentication behavior, settlement correctness, refunds, reconciliation, exception handling, and rollback paths.

05 · READOUT

Measure creator access and payment reliability together.

49%Creators saying late or inconsistent payments affected business operations
41%Creators saying cash flow issues forced them to turn down opportunities
62%Creator businesses using personal bank accounts for business transactions
06 · LEARNING

Creator payments are a product capability, not only infrastructure.

The strongest experience makes earnings more usable while keeping money movement safe and operationally correct. The user experience can stay coherent even when the underlying rails vary by market.

07 · TRADEOFFS & RISK

I chose operational correctness over a faster, uniform rollout.

The tradeoff was launch speed versus market-specific reliability. Reusing one payment flow everywhere would be simpler, but authentication, settlement, refunds, compliance, and partner behavior vary by market. I accepted more integration complexity and a slower rollout sequence so the creator experience could remain consistent without hiding fragile money movement underneath.

Risk acceptedMore implementation overhead in exchange for fewer settlement failures, cleaner rollback paths, and safer expansion.
Key decisions & metrics

Improve creator access to earnings without compromising payment correctness, risk controls, or cost to serve.

Key decisionMake cash flow part of the creator product

Treat payout timing, spendability, and business money management as a creator experience instead of invisible payment infrastructure.

Primary outcomeFaster access to usable earnings

Measure time from earned rewards to spendable funds, successful card activation, successful spend rate, payout completion, and creator support contacts.

Time to usable fundsActivationSuccessful spendSupport contacts
GuardrailsPayment correctness and risk

Track fraud and chargebacks alongside settlement accuracy, refunds, reconciliation, authentication failures, and cost per successful transaction.

Fraud / chargebackSettlement accuracyRefund successCost / transaction
Scale gateMarket readiness before expansion

Expand only after partner readiness, regulatory checks, golden transactions, settlement, refunds, exception handling, and rollback behavior all pass.

01 · Problem

Rare defects, scarce labels, low trust

Traditional supervised inspection needs many labeled failures, which are exactly the samples factories rarely have. The product problem was to detect and localize unusual defects using predominantly normal imagery, while keeping false alarms low enough for operators to trust the system.

02 · What I did

Turned model research into an operational workflow

  • Defined alert quality and false-positive reduction as the North Star instead of raw alert volume.
  • Structured the path from image acquisition and dataset configuration through model benchmarking, threshold calibration and human review.
  • Designed explainable outputs such as anomaly score, heat map, prediction mask and segmentation overlay, so engineers could see why a part was flagged.
  • Partnered across research, engineering and users to close the loop between model quality, edge performance and adoption.
03 · What we used

Modular computer-vision stack

AnomalibPyTorch LightningPaDiMPatchCoreFastFlowOpenVINOPython API + CLIYAML configsCPU / edge inferenceAnomaly mapsSegmentation masksBenchmarking
04 · Impact

Better quality, efficiency and adoption

-15%false positives
+20%inference efficiency
+14%platform MAU

The key product result was not simply detecting more anomalies. It was giving engineers evidence they could inspect, tune and trust.

1AcquireMostly normal production images
→
2BenchmarkPaDiM · PatchCore · FastFlow
→
3CalibrateScores, thresholds and callbacks
→
4DeployExport and run with OpenVINO
→
5ExplainHeat map · mask · segmentation
Official Anomalib model output showing the source image, anomaly map, prediction mask and localized defect
Model output · localized anomalyThe inference result connects the source image to its anomaly map, mask and localized defect so engineers can inspect why it was flagged.
Anomalous cube dataset example with a simulated black circular manufacturing defect
Anomaly dataset · simulated defectA black circular marker simulates a hole or manufacturing defect, creating a clear abnormal sample for validation and inference.
Open-source contribution · visual anomaly detection · edge AI
Key decisions & metrics

For anomaly detection, model quality must be balanced with false alarms, explainability, edge performance, and operator trust.

Key decisionChoose the model against the deployment envelope

Benchmark PaDiM, PatchCore, and FastFlow on detection quality together with latency, memory footprint, calibration stability, and edge deployability. The best research score is not automatically the best product choice.

North StarActionable alert quality

False positive reduction is a strong product outcome because noisy alerts destroy trust. Model metrics should support that outcome, not replace it.

-15% false positives+20% inference efficiency+14% platform MAU
Model + system metricsMeasure detection, localization, and serving

Track image level AUROC or F1, pixel level localization quality, p95 inference latency, memory footprint, threshold drift, and operator review time.

Image AUROC / F1Pixel localizationp95 latencyThreshold drift
Scale gateRequire trustworthy explanations

Do not broaden deployment if gains come from excessive alerting or if heat maps and masks are too unstable for engineers to validate.

01 · PROBLEM

Risk teams need one view of decisions, evidence, and model health.

Credit and fraud workflows often fragment streaming signals, model outputs, investigations, policies, and monitoring across separate tools. That makes decisions slower to explain and harder to operate consistently.

02 · PRODUCT DECISION

Treat risk as a decision platform, not a single prediction model.

I designed the product around the full analyst workflow. Models score risk, anomaly logic surfaces unusual behavior, explanations expose the contributing signals, and policy retrieval gives analysts the context needed to act.

03 · SYSTEM DESIGN

Separate streaming, feature processing, inference, orchestration, and retrieval.

Kafka handles event ingestion, PySpark transforms transaction data, TensorFlow produces model scores, MLflow manages model lifecycle, Airflow coordinates recurring workflows, and the RAG layer retrieves relevant policy guidance.

04 · HUMAN WORKFLOW

Keep analysts in control of consequential decisions.

The product surfaces scores, anomalies, explanations, and evidence together so an analyst can investigate instead of blindly accepting an automated decision. The system is designed to support judgment and escalation.

05 · OPERABILITY

Monitor the model and the decision process.

The platform includes model monitoring and workflow observability so teams can inspect score behavior, data quality, orchestration health, and investigation patterns as the system evolves.

06 · PRODUCT TAKEAWAY

Trust comes from traceability across the whole system.

A useful risk product must connect prediction quality with evidence, policy, human review, and operational monitoring. The model is one component of the product, not the product itself.

Key decisions & metrics

Fraud and credit are related risk domains, but they should not share one blended quality metric.

Key decisionSeparate fraud effectiveness from credit decision quality

Fraud optimizes harmful behavior detection with manageable false positives. Credit decisioning needs discrimination, calibration, approval trade offs, and loss performance.

Fraud scorecardCatch true risk without flooding analysts

Track recall at fixed precision, false positive rate, confirmed fraud catch rate, alert to case conversion, time to detect, and loss prevented. Add values only after a real evaluation run.

Recall @ fixed precisionFalse positive rateTime to detectLoss prevented
Credit scorecardMeasure ranking and calibration separately

Track AUC or KS, probability calibration, approval rate under a risk constraint, default or loss rate, and fairness diagnostics where appropriate.

AUC / KSCalibrationApproval rateLoss rate
Operating guardrailsMake the decision system observable

Track analyst review SLA, cost per reviewed alert, data and feature drift, model drift, p95 inference latency, pipeline health, and override patterns.

Review SLACost / reviewed alertDriftp95 latency
01 · DIAGNOSIS

The pattern pointed to a system problem, not one bad batch.

Quality was degrading across multiple dimensions at the same time. That suggested a breakdown in policy clarity, calibration, vendor consistency, and edge case handling rather than an isolated labeling miss.

79.8%Good driving rate in late batches, down from roughly 87 to 91%
68 to 73%Audit agreement in late batches, down from roughly 85 to 90%
6×Increase in rework from January to February and March
84 to 86%Late vendor consistency, down from roughly 93%
02 · NORTH STAR

Weekly Trusted Label Acceptance Rate

The proposed North Star asks one useful product question: what share of labeled segments can move into model training confidently, without rework?

Accepted segments that pass audit with no rework÷Total labeled segments

The 87% example in the source deck is illustrative, not a claimed production result.

03 · FIRST 30 DAYS

Repair the quality operating system first.

01Clarify labeling rulesTurn broad concepts such as “good driving” into explicit decision criteria.
02Build an edge case handbookCreate shared examples for construction zones, crosswalks, yielding, visibility, hesitation, and avoidability.
03Run weekly quality reviewsTrack the North Star alongside diagnostic metrics and recurring error types.
04Calibrate labelers and vendorsUse disagreement cases to align interpretation before more data enters training.
05Create an escalation loopFeed unresolved ambiguity back into policy and the handbook.
04 · PRIORITIZATION

Foundation first. AI second.

An AI copilot is not the first move when policy is unclear and drift is poorly observed. The stronger sequence is to establish shared policy memory and quality visibility, then add automation where it can make consistent decisions.

01Closed Loop Policy Learning SystemCapture resolved edge cases into versioned policy.
02Edge Case Knowledge Graph and HandbookCreate a shared source of truth across labelers, QA, and vendors.
03Real Time Drift and Anomaly DetectionSurface quality deterioration before it contaminates training runs.
04Guided Labeling Decision EngineStandardize judgment once the policy foundation is strong enough.
05 · PRODUCT VISION

Driving Quality Intelligence Platform

Guided Decision System→Continuous QA Loop→Model Label Feedback→Explainability Layer

The future state turns labeling quality into a closed loop product system. Decisions become structured, resolved edge cases update policy, drift becomes visible earlier, and every label can be traced back to the reasoning behind it.

06 · PRODUCT TAKEAWAY

Training data quality is a product problem.

The value is not simply faster annotation. The product must create reliable judgment across policy, calibration, shared case memory, observability, escalation, and automation. AI becomes useful after that foundation is trustworthy.

Key decisions & metrics

The measurement system should connect labeling quality to model readiness without letting throughput or cost hide degraded judgment.

Key decisionFix policy and calibration before an AI copilot

When labelers disagree on the definition of good driving, automation scales inconsistency. Build shared policy memory, edge case examples, calibration, and quality visibility first.

North StarWeekly Trusted Label Acceptance Rate

Measure the share of labeled segments that pass audit and can enter model training without rework. This links labeling operations directly to downstream model readiness.

Diagnostic metricsSeparate quality, consistency, operations, and coverage

Track audit accuracy, inter rater agreement, rework, SLA attainment, average label time, vendor consistency, complex scene coverage, critical error miss rate, escalation rate, and cost per accepted label.

Audit agreementInter rater agreementReworkSLACost / accepted label
Scale gateDo not trade quality for throughput

Faster labeling is not a win if trusted acceptance falls or downstream model regressions increase. Stop rollout when critical misses, disagreement, or model impact breach thresholds.

1 / 1

Product studio archive

Earlier product, GTM (go-to-market) and analytics work - where the discovery, experimentation and data-storytelling muscles got built.

Marketing and product strategy animation
Product studio archive · evidence, not thumbnailsEach project now carries its original visual artifact, from research and GTM work to dashboards, experiments, and shipped product screens.
INGRAM MICRO · INDUSTRYB2B mobile app & payment rails

Shipped a B2B mobile app to 100K MAU and launched ACH, SEPA and SWIFT rails - contributing to a 22% reduction in payment fraud along the way.

Full history on LinkedIn →
Money App transaction-history screen
History
Money App transfer form screen
Transfer
Money App favorite-friends transfer screen
Recipients
2022 · FINTECHMoney App - digital wallet 0 → 1

Built a digital money transfer wallet with engineering and shipped the GTM strategy; the wallet became a meaningful revenue contributor.

Product screens included above →
Google Photos engagement projectRetention strategy
2022 · ENGAGEMENTGoogle Photos - retention strategy

User personas, RICE-prioritized pain points, and an Agile roadmap targeting retention, churn and success metrics.

View deck →
Discord product strategy visualGTM expansion
2021 · GTM STRATEGYDiscord - beyond gaming

Feature strategy and GTM plan to expand Discord past gaming communities, reach new segments, and shift brand perception.

View deck →
Unilever consumer research visualConsumer research
2021 · CONSUMER RESEARCHUnilever SE Asia - CMI research

Consumer behavior research with Qualtrics, A/B test analysis, and a SWOT-based recommendation presented to the Consumer & Market Insights team.

View presentations →
Spotify Tableau dashboardTableau dashboard
2021 · DATA STORYTELLINGSpotify - global growth dashboard

Tableau deep-dive on growth trends, demographics, market share, revenue and competitors. More dashboards on my Tableau profile.

View dashboard →
A/B testing comparisonExperiment design
2022 · EXPERIMENTATIONSurf - creative A/B testing

Split-tested print and digital creative to isolate which variables drove the most traffic and clicks.

View projects →
Denny's customer segmentation visualR · clustering
2021 · ANALYTICSDenny's vs IHOP - segmentation in R

Survey-based customer segmentation with clustering, dendrograms and hypothesis testing - the analytical foundation under the PM work.

See the code →
1 / 1

Engineering projects, built hands-on

Hands-on software builds behind the product thinking: connected devices, web applications, APIs, utilities, and interactive prototypes.

Connected-home automation overviewConnected home
IoT · ARDUINO / NODEMCUHome automation system

WLAN and RF control for home devices with manual, automatic, and security modes.

Explore GitHub →
Android Studio prototype of the Trip Adviser applicationAndroid prototype
NODE.JS · HANDLEBARSTrip Adviser

A trip-creation web application with dedicated chat rooms for travelers.

View code →
Sublease Finder application running in an Android emulatorMobile marketplace
NODE.JS · PAYMENTSSublease Finder

Sublease discovery with in-app messaging and payment-flow concepts.

Explore GitHub →
Two-player browser dice game interfaceJavaScript game
JAVASCRIPTTwo-player Dice Game

A browser-based game demonstrating event handling, state updates, and responsive interaction.

View code →
People Finder Express.js application search interfaceExpress · JSON search
EXPRESS.JS · JSONPeople Finder

A search application for filtering and retrieving people from structured JSON data.

View code →
Prime Number Checker Node.js application interfaceNode.js utility
NODE.JSPrime Number checker

A focused Node.js utility for validating numbers and handling user input.

View code →
GitHub Finder React application displaying profile search resultsReact · GitHub API
REACT · GITHUB APIGitHub Finder

A React application for searching GitHub profiles and displaying public account data.

View code →
1 / 1
About

I build AI products from ambiguity to adoption.

As a Staff Product Manager with 8+ years across engineering and product leadership, I can enter a new problem space, learn the system quickly, and lead the full journey from discovery and product strategy through model behavior, evaluation, launch, and scale.

Product judgment with technical depth

I’m a customer-obsessed PM who turns ambiguous ideas into products people actually use. I start with the customer problem, connect it to business value and technical reality, and stay accountable for what we build, why it matters, and how we know it works.

My strength is connecting the layers that make AI products successful: user needs, data quality, model capability, workflow design, economics, and operational reality. I partner deeply with engineering, data science, design, research, and GTM while making the tradeoffs explicit across quality, latency, cost, guardrails, and human fallback.

I do not default to a single domain playbook or to AI itself. I learn the customer, incentives, constraints, data, and failure modes first, then choose the simplest product system that fits the problem, ship it in stages, measure real value, and scale what works.

Customer problem first Model + data fluency Evals before scale Human-in-the-loop 0-to-1 through growth Outcome ownership
A strong AI product is not a model wrapped in a UI. It is a reliable system of data, decisions, workflows, feedback, and measurable value.
Tanishq Mishra at graduation
University of San FranciscoMS in Marketing Intelligence · Product Management2022
AI product leadership

AI Product Manager ≠ Product Manager who uses AI

Both matter. The difference is whether AI improves the PM's workflow, or whether the PM owns the intelligence, risks, and economics of the product itself.

AI-enabled workflow

A Product Manager who uses AI

Uses tools such as ChatGPT, Claude, and Cursor to move faster.

  • Draft and synthesize faster
  • Analyze customer feedback
  • Generate and test ideas
  • Automate repetitive work
AI as the product

An AI Product Manager

Builds products where AI is the core capability and owns the hard product decisions.

  • Choose the model, data, retrieval, and tool architecture
  • Define quality, evaluation, and launch thresholds
  • Reduce hallucinations and design safe failure paths
  • Balance accuracy, latency, reliability, and cost

The strongest modern PMs should be able to do both. I use AI to improve how I work, and understand how to build, evaluate, launch, and scale AI-powered products.

Contact

Building an ambitious AI product?

I'm in the SF Bay Area and enjoy working with teams that are turning complex technology into useful, trusted, and scalable products, from the first hypothesis to measurable adoption.

Tanishq0630@gmail.com