Tushar Ghosh / IIIT Nagpur / 2027
Reliable automation. Useful products. Trustworthy data.
I build AI automation that connects triggers, agents, integrations and human review, alongside full-stack applications and analytics pipelines with persistent state and thoughtful failure handling.
React · Node.js · Python · SQL · PostgreSQL · Airflow / dbt · Power BI · LangGraph
Generative AI & Machine Learning Engineer Intern at MyManah ↗ / Remote / Oct 2026 - Present
01 / Experience
IIIT Nagpur / B.Tech CSE - Data Science & Analytics / Expected 2027
Generative AI & Machine Learning Engineer Intern
Python / LangGraph / FastAPI / Hybrid RAG / Pytest
Employer and external contributions, excluding my personal projects. Private PR titles and implementation details remain private.
02 / AI automation & agentic workflows
Research, approvals, campaigns and payment recovery, with explicit policies, persistent state and inspectable handoffs.
An RSS event triggers Zapier intake, CrewAI research and AutoGen critique. A version-bound human decision gates Slack delivery; Neon checkpoints and a transactional outbox preserve the run.
Expense intake flows through deterministic policy, durable n8n approval waits and optional Slack/email delivery. AI explains and investigates; financial decisions remain with explicit rules and human reviewers.
An n8n/Gemini agent retrieves scoped inquiry and service knowledge, then proposes a cited qualification and follow-up draft. Human review checks the evidence; approval atomically creates a task and audit record, not an outbound message.
Hosted CRM requires a workspace account. Live AI depends on local n8n and a webhook tunnel.
A travel request routes through five specialists and typed provider evidence, then pauses for human itinerary review. PostgreSQL checkpoints support interrupt/resume, and approved plans become session-bound PDFs. Research and planning, not booking.
Payment recovery automation / Public product
Signed Razorpay failure webhooks create persisted recovery cases and Redis-worker tasks. ML triage and LangGraph propose retry, defer or stop strategies; deterministic guardrails and high-value human approval gate actions and preserve an audit trail.
* Offline evaluation on 523 held-out synthetic events. Demo payment flows use test mode; no production recovery claim.
These are independent portfolio projects. Deployment, simulation and reference-build boundaries are stated per project; they are not claims about employer systems.
03 / Full-stack & data platforms
Full-stack workspaces, deployed products, tested analytics models and forecasting.
A full-stack AI workspace with an Express gateway, four Node.js services, MongoDB and Redis, and LangGraph routing across eight specialist paths.
A deployed React and Express product for building voice assistants that site owners can embed on third-party websites.
Replay-safe PostgreSQL CDC through Airflow and Spark/Delta into dbt silver and gold models, with SCD history, grain checks and a query-plan lab.
NYC taxi analytics, leakage-safe zone-hour forecasting and route intelligence, connected to a dashboard, API and constrained fleet-allocation simulation.
04 / Data analytics & engineering
Business questions, SQL models, data quality and dashboards, with results connected to their analytical context.
Azure-ready JSON ingestion and PySpark transformation into a star schema, with Power BI sales analysis and targets-versus-actuals reporting.
Vendor-brand SQL aggregation, statistical testing and Power BI reporting connect purchasing, margin and inventory to procurement decisions.
A React and Flask e-commerce dashboard backed by Pandas ETL, PostgreSQL and Redis-cached KPIs, with CSV exports and inspectable ETL run history.
Retail reports use simulated/public data. Vendor findings describe the analyzed population and do not establish causal effects.
05 / Fine-tuning, vision & speech
Vision-language adaptation and speech annotation: trained adapters, pretrained pipelines and inspectable evaluation evidence.
Fine-tuned Qwen3-VL-2B with NF4 QLoRA on a Tesla T4 to extract merchant, address, date and total from receipt images. Merchant-disjoint validation, content audits and saved base-versus-adapter outputs make the experiment inspectable.
239-image retrospective comparison, not a fresh-domain benchmark. Latest Qwen inference uses a verified CLI; the API still serves legacy SmolVLM V2.
Combines pyannote speaker diarization, wav2vec2 speech-emotion classification and Whisper language identification into per-segment annotations. Built Streamlit/Plotly and batch CLI workflows with cached models, inference locks and reusable WAV/RTTM artifacts.
Pretrained model integration, not speech fine-tuning. Emotion labels are model predictions; validated emotion accuracy and production throughput are not claimed.
06 / How I engineer
I use Codex, Claude and Gemini for repository exploration, implementation, debugging and repetitive changes. I verify their output through tests, CI, logs, traces and the running system. Architecture and product decisions remain mine.
07 / More work
Stateful research, local inference, retrieval, multimodal ML and operational analytics.
LeadDesk reviewed on 1 November 2026; remaining automation summaries were reviewed on 23 October 2026. Other project evidence retains its previously documented verification scope.
08 / Problem solving
1,250+ DSA problems solved across coding platforms.
09 / Contact
Currently engineering agent systems at MyManah. Interested in thoughtful conversations about AI automation, full-stack products and trustworthy data.