Projects

Featured projects

01 / Flagship system

ConceptGraph AI

Course-aware GraphRAG that turns PDFs into searchable concept graphs, reranks evidence, and returns page-grounded answers.

ReactFastAPIPostgreSQLQdrantNeo4j

Inside the build

PDF extraction / bounded OCR
Page-aware chunks → PostgreSQL
Embeddings → Qdrant
Concept extraction → Neo4j
READY-gated vector candidates
Graph relationship expansion
Reranking
Grounded answer

Key decision

Immutable course UUIDs align records across stores. PostgreSQL-backed leases and attempt fencing make a bounded, in-process worker restart-safe without adding a separate queue service. READY gating prevents partially ingested documents from entering retrieval; page and chunk provenance survive expansion and reranking.

What needed care

An accepted upload is not a usable document. Partial graph extraction, provider rate limits, interrupted jobs, and cross-store deletion all need explicit states. Resumable section checkpoints, provider failover, attempt fencing, and upload-scoped cleanup make these failure paths inspectable.

22 Questions · retrieval ablation

Case study

Agent engineering / support operations

In progress

Relay

A support-investigation prototype that plans, retrieves evidence, verifies its proposal, and waits for human approval before creating a review ticket.

PythonSQLiteBM25Gemini/OpenAI APIsJavaScript

Inside the build

Issue → persistent SQLite queue
Worker records investigation plan
Model selects scoped tools
BM25 policies + account/service fixtures
Evidence → plan revision
Verify exact proposal
Human approves or rejects
Idempotent local review ticket

Key decision

A bounded single-agent loop uses Gemini/OpenAI adapters and explicit tools rather than arbitrary code execution. Verification is bound to the exact proposal and is invalidated when evidence or the plan changes. SQLite transactions enforce one local ticket per approved run.

What needed care

Demo scenarios use fictional accounts, policies, and incidents. Scripted mode is deterministic; live mode uses model-selected tool calls. Mechanical evidence checks do not establish semantic correctness or prompt-injection resistance. The local server has no production authentication or tenant isolation, and approval creates a local ticket rather than a refund or an external support action.

42 Tests · documented local suite

Case study

Distributed systems / job execution

In progress

FaultLab

A PostgreSQL-backed job runner that makes retries, worker crashes, lease recovery, and cancellation visible—with Python and Go workers.

PythonGoFastAPIPostgreSQLDocker Compose

Inside the build

FastAPI validates and accepts jobs
PostgreSQL stores jobs and attempts
Workers claim with SKIP LOCKED
Leases and heartbeats fence attempts
Registered handlers execute work
Retry, cancel, or recover expired leases

Key decision

PostgreSQL is the coordination layer: short transactions claim work atomically, while Python and Go processes share a lease protocol. Submission idempotency is separate from business-effect idempotency. Execution is at least once; arbitrary external side effects are not exactly once.

What needed care

Cancellation is cooperative, so a handler that ignores it can still succeed. SKIP LOCKED does not guarantee global fairness. The lab has no tenant isolation, arbitrary-code sandbox, or validated backup strategy. Local benchmark observations are not production capacity or service-level guarantees.

2 Worker runtimes · Python / Go

Case study

Machine learning / race simulation

F1 Race Strategy Decision Support

Lap-time prediction and Monte Carlo simulation for comparing pit-stop strategies using real 2021–2025 F1 data.

PythonStreamlitscikit-learnPandasNumPy

Inside the build

FastF1 telemetry
Feature engineering
Rolling split
Lap-time model
Monte Carlo simulation
Strategy bands
Interactive dashboard

Key decision

Use the latest available earlier-season median for each circuit as the pre-race baseline, excluding the completed test race. Compare Ridge with HistGradientBoosting and sample residuals plus pit loss to return P10/P50/P90 strategy-time bands.

What needed care

Wet-weather laps and Las Vegas produce larger errors. The model is more trustworthy in dry-race scenarios and does not yet explicitly simulate traffic, safety cars, or overtakes.

1.751s HGB MAE · held-out 2024 laps

Case study

Privacy-aware wellbeing product

Serenity AI

A mood-aware journaling and conversation product designed around secure personal data, recent context, and restrained AI assistance.

ReactTypeScriptViteSupabasePostgreSQL

Inside the build

Mood check-in
Authenticated write
Recent context
Gemini response
Wellness insight

Key decision

User-scoped tables are protected with Row Level Security, and model context is deliberately bounded to the latest six messages.

What needed care

Optimistic updates initially made failed writes too quiet. The interface needed an explicit pending state and a reversible correction path.

6 Message context bound

Case study

Browser-first product

Résumé ATS Analyzer

Parses résumés locally, scores five ATS dimensions, compares job descriptions, and exports actionable feedback.

ReactTypeScriptViteTailwind CSSRadix UI

Inside the build

PDF / DOCX / TXT
Browser parsing
Five-dimension scoring
Job keyword comparison
Feedback report
HTML export

Key decision

Core parsing and scoring run locally. Weighted checks provide explainable feedback rather than presenting an opaque score as a hiring prediction. A local history keeps the ten most recent analyses.

What needed care

Complex document layouts can scramble extracted text, while keyword matches lack the judgment of a recruiter. Scores are heuristic guidance, not a guarantee of compatibility with every employer's ATS or of an interview.

5 Explainable scoring dimensions

Case study

From the workbench.

Smart Classroom Scheduler

A timetable-planning concept. The public repository currently contains a README and license, not an implemented scheduler.

Concept / documentation

Repository

Mutual Fund Analytics

Internship work at Bluestock: analysis across 40 schemes, 10 datasets, 87K+ records, and 4.5 years of NAV history.

Internship work / no public repository

Phishing Website Detection

A PhiUSIIL dataset investigation covering preprocessing, baseline models, Optuna tuning, and saved inference artifacts.

Python / machine learning

Repository

Featured projects

01 / Flagship system

ConceptGraph AI

Course-aware GraphRAG that turns PDFs into searchable concept graphs, reranks evidence, and returns page-grounded answers.

ReactFastAPIPostgreSQLQdrantNeo4j

Inside the build

PDF extraction / bounded OCR
Page-aware chunks → PostgreSQL
Embeddings → Qdrant
Concept extraction → Neo4j
READY-gated vector candidates
Graph relationship expansion
Reranking
Grounded answer

Key decision

What needed care

22 Questions · retrieval ablation

Case study

Agent engineering / support operations

In progress

Relay

A support-investigation prototype that plans, retrieves evidence, verifies its proposal, and waits for human approval before creating a review ticket.

PythonSQLiteBM25Gemini/OpenAI APIsJavaScript

Inside the build

Issue → persistent SQLite queue
Worker records investigation plan
Model selects scoped tools
BM25 policies + account/service fixtures
Evidence → plan revision
Verify exact proposal
Human approves or rejects
Idempotent local review ticket

Key decision

What needed care

42 Tests · documented local suite

Case study

Distributed systems / job execution

In progress

FaultLab

A PostgreSQL-backed job runner that makes retries, worker crashes, lease recovery, and cancellation visible—with Python and Go workers.

PythonGoFastAPIPostgreSQLDocker Compose

Inside the build

FastAPI validates and accepts jobs
PostgreSQL stores jobs and attempts
Workers claim with SKIP LOCKED
Leases and heartbeats fence attempts
Registered handlers execute work
Retry, cancel, or recover expired leases

Key decision

What needed care

2 Worker runtimes · Python / Go

Case study

Machine learning / race simulation

F1 Race Strategy Decision Support

Lap-time prediction and Monte Carlo simulation for comparing pit-stop strategies using real 2021–2025 F1 data.

PythonStreamlitscikit-learnPandasNumPy

Inside the build

FastF1 telemetry
Feature engineering
Rolling split
Lap-time model
Monte Carlo simulation
Strategy bands
Interactive dashboard

Key decision

What needed care

Wet-weather laps and Las Vegas produce larger errors. The model is more trustworthy in dry-race scenarios and does not yet explicitly simulate traffic, safety cars, or overtakes.

1.751s HGB MAE · held-out 2024 laps

Case study

Privacy-aware wellbeing product

Serenity AI

A mood-aware journaling and conversation product designed around secure personal data, recent context, and restrained AI assistance.

ReactTypeScriptViteSupabasePostgreSQL

Inside the build

Mood check-in
Authenticated write
Recent context
Gemini response
Wellness insight

Key decision

User-scoped tables are protected with Row Level Security, and model context is deliberately bounded to the latest six messages.

What needed care

Optimistic updates initially made failed writes too quiet. The interface needed an explicit pending state and a reversible correction path.

6 Message context bound

Case study

Browser-first product

Résumé ATS Analyzer

Parses résumés locally, scores five ATS dimensions, compares job descriptions, and exports actionable feedback.

ReactTypeScriptViteTailwind CSSRadix UI

Inside the build

PDF / DOCX / TXT
Browser parsing
Five-dimension scoring
Job keyword comparison
Feedback report
HTML export

Key decision

Core parsing and scoring run locally. Weighted checks provide explainable feedback rather than presenting an opaque score as a hiring prediction. A local history keeps the ten most recent analyses.

What needed care

5 Explainable scoring dimensions

Case study