Full Stack Engineer · AI Product Builder · Pune

I build products from idea to production.

0→1 product development across backend systems, APIs, AI workflows and automation — with a bias toward shipping and a habit of designing for failure.

3live products shipped
2,000+GAINN signups
21/21RecoverIT validation checks
40%less manual testing · Ceeras
System state

Interpret → structure → verify → act.

working pathadaptive
RESEARCH / QC
AI + DETERMINISTIC
SHIP / OBSERVE
system thinkingproblem → signal → workflow
3 live productsend-to-end shipping
APIs + workflowsbackend / automation
21 / 21RecoverIT test-mode checks
2,000+GAINN signups
40%less manual effort
01 / Selected work

Serious systems.
Less decoration. More evidence.

Four flagship builds showing how I move from an ambiguous problem to a working product: architecture, APIs, AI workflows, reliability controls and user-facing delivery.

LiveAI + deterministicPayments21 / 21 test-mode checks

RecoverIT

An AI-assisted payment recovery workflow for ambiguous failed-payment cases.

AI interprets messy evidence. Deterministic workflow state controls consequential actions. Humans approve high-risk changes.

case ledgerreliabilitysafety boundary
Core boundaryAI interpretation → durable case + workflow ledger → deterministic rules → action.
Validation21 / 21 checks passed in synthetic / test mode; not production recovery data.
LIVE PRODUCT ↗
Actual RecoverIT live product interface
ACTUAL PRODUCT UI · RECOVERIT
Live · 2,000+ signupsMulti-agentInformation intelligence

GAINN

From information overload to understanding.

A multi-agent news pipeline that separates drafting from quality control so publishing is not one opaque generation pass.

researchdraftingQC
FlowRaw information → research → draft → QC → context.
Why it existsGeneration is not usefulness; verification needs a separate responsibility boundary.
LIVE PRODUCT ↗
Actual GAINN live product interface
ACTUAL PRODUCT UI · GAINN
Live · 1,000+ signupsResearch pipelineVideo

StoryLens

From scattered research to a coherent story.

A research-to-video pipeline that separates retrieval, cross-verification, entities, narrative synthesis and animated output.

retrievalverificationnarrative
FlowResearch → sources → cross-check → entities → narrative → script → video.
Key insightResearch output needs verification and structure before it becomes useful narrative.
LIVE PRODUCT ↗
Actual StoryLens live product interface
ACTUAL PRODUCT UI · STORYLENS
Research prototypeTelemetry observabilitySynthetic data30 tests · 95.68% coverage

FleetScope

A telemetry reliability lab for finding delivery and data-quality issues before they become misleading conclusions.

A reproducible fleet-event pipeline that measures freshness, retries, conflicting replays, unchanged readings, offset continuity and expected-signal cadence — with explicit safeguards against claiming more than the data proves.

Python · FastAPISQLitePrometheusSQL pagination
PipelineDeterministic simulator → strict event validation → batch analysis → SQLite history → FastAPI, dashboard, CSV and Prometheus metrics.
SignalsFreshness percentiles, retries, replay conflicts, unchanged values, offset gaps and reordering, plus estimated missing emissions against declared cadence.
Verification30 local tests at 95.68% line coverage, real Uvicorn smoke checks, and a documented 50,000-base-event scale check.
BoundariesAll data is synthetic. Offset gaps are not proof of data loss; cadence gaps are estimates. This is a batch-oriented independent prototype, not a Tesla system or production SLO.
FLEETSCOPE / TELEMETRY LABSYNTHETIC
Fleet health / balanced profile
Events993accepted
Freshness p956.6ssample run
Resends33identified
SIGNAL QUALITY · SAMPLE RUN
Late events
88
Future time
9
Offset gaps
8
Reordering
3
SEED 23 · 12 SYNTHETIC VEHICLESANALYSIS COMPLETE
FLEETSCOPE · TELEMETRY RELIABILITY
Engineering lab
01
RecoverIT failure mode
AI is available. The workflow continues through evidence, investigation and safe next action.
workflow continuesRECOVERABLE
02
GAINN QC
QC pass unlocks publishing; the pipeline keeps the responsibility boundary explicit.
publishSTATE CHANGES
03
StoryLens source conflict
Sources agree, so the system can move from verification into narrative structure.
continue to narrativeVERIFY PATH
02 / Experience

Research when the problem is new.
Engineering when it has to work.

A compact timeline: two environments, two kinds of ambiguity, one habit of learning the system underneath the task.

DEC 2025 — JAN 2026
Research Intern
IDSIA / SUPSI · Supervisor: Paolo Franceschi

Worked on training-data quality for imitation learning and Vision-Language-Action (VLA) models.

  • Data collection and augmentation for robot-learning pipelines.
  • Multimodal learning-from-demonstration data quality and diversity.
FEB 2025 — MAY 2025
Penetration Tester
Ceeras Pvt. Ltd.

Conducted security assessments across internal and web systems.

  • Automated repeatable testing workflows, reducing manual effort by 40%.
  • Worked across application and security assessment tasks.
03 / Other work

Different problems.
Same systems instinct.

LLM safety evaluator

Red-team → judge → report

Automated evaluation across GPT, Claude, DeepSeek and Gemini with LLM-as-a-Judge scoring, FastAPI, Streamlit, LiteLLM and PDF reporting.

GPTCLAUDEDEEPSEEKGEMINILLM-AS-A-JUDGE85% DETECTION
LHC anomaly detection

100k+ simulated events → anomalies

Unsupervised Autoencoder + Isolation Forest across a simulated LHC-style dataset, with a React / TypeScript dashboard for inspection.

100K+ SIMULATEDAUTOENCODERISOLATION FORESTREACT / TS
Human–AI interaction research

Beyond the Chat Box

Ten exploratory essays on multimodal interaction, voice and silence, agent autonomy, memory, multilingual interfaces, ambient AI, social agents and human capability.

10 ESSAYSHCIMULTIMODALAGENTIC SYSTEMS
04 / Skills

A broad stack.
Used with intent.

The point is not the list. It is how the layers connect: research → model → API → workflow → product.

Languages

Primary implementation tools
PythonTypeScriptJavaScriptC++JavaBash

AI / ML

Models, evaluation and orchestration
PyTorchscikit-learnLLM-as-a-JudgeMulti-agent systemsAnomaly detectionAdversarial prompt testing

Backend / Full Stack

Application, APIs and infrastructure
Node.jsExpressFastAPIREST APIsReactMongoDBMongooseDockerLinuxGitLiteLLM

Data

Analysis and pipelines
NumPypandasStatistical modelingData pipelines

The stack, in motion.

Move through the layers. The same principle appears across the projects: turn uncertain inputs into something a deterministic system can use.

WORKFLOWresearch → build → verify
researchmodelAPIproduct
05 / Achievements

Signals that earned attention.

HackathonTop 50 / ~250k

ET GenAI Hackathon — approximately top 0.02% of participants.

ProgramHPAIR

Selected delegate, Harvard Project for Asian and International Relations.

Education9.74 / 10

B.Tech, Information Technology · Bharati Vidyapeeth College of Engineering, Pune.

ProfileTechnical Generalist

AI systems, product engineering, research, automation and security.

06 / About

A full-stack engineer with a product and systems bias.

Prakash Chand Jain
B.Tech IT
BVCOE Pune
GPA
9.74 / 10

I’m Prakash Chand Jain, a full-stack engineer and AI product builder who likes taking ambiguous ideas and turning them into working systems.

My work crosses backend engineering, AI orchestration, automation, research and security. I care about the layer underneath the interface: APIs, data, state, failure modes, recovery, verification and what keeps working after launch.

01Start with the problem
02Question defaults
03Find the abstraction
04Build the useful core
05Test failures
06Ship and iterate
07 / Contact

Need someone to take a product from idea to launch?

I’m strongest where there is no clean playbook yet: shaping the product, choosing the architecture, wiring the integrations, building the workflow and shipping the useful version quickly.