Sanskar Gupta — AI/ML & Generative AI Engineer

03 — Projects

Agents, pipelines, and tools.

FinPilot — Autonomous AI CFO

Jul — Sep 2026

Autonomous financial copilot with guardrail-first orchestration — 5 agents, LangGraph, JWT auth

Created an autonomous financial copilot powered by 5 specialized LangGraph agents across a Node.js/Express + FastAPI + PostgreSQL stack for cash flow, runway, and forecast analysis grounded in live transaction data — not a hardcoded balance.
Designed a guardrail-first orchestration layer with JWT-based auth, prompt-injection input sanitization, and Pydantic output-schema validation, all wired in before any agent or LLM call — a blocked question never touches the database.
Enabled a conversational CFO Chat agent to replace static dashboards with explainable, real-time insights, showing its reasoning instead of hiding it behind a single generated answer.
Backtested the forecasting model against real held-out data and validated anomaly detection with explainable z-score statistics rather than a black-box model, with an 18-question golden eval set and 31 passing unit/integration tests.

PythonFastAPILangGraph Node.js / ExpressPostgreSQLJWT StreamlitDocker

Architecture — one clear entry point, reads and writes split by service

Streamlit (chat + dashboard) → Node.js / Express (API layer + proxy · owns writes) → FastAPI + LangGraph (5 agents · owns reads-for-reasoning) → PostgreSQL

Agents — one file, one responsibility each

Agent	Responsibility	Key tech
Cashflow Agent	Computes inflows, outflows, burn rate, and spend by category from real transaction history	Pandas, SQL aggregation
Runway Agent	Projects months of cash remaining, grounded in actual transaction data	PostgreSQL, SQLAlchemy
Forecast Agent	Produces a transparent linear-trend forecast, backtested against held-out data with reported mean absolute error	NumPy statistics
Risk Agent	Flags unusual spending and missing/delayed revenue using explainable z-score statistics	Statistical anomaly detection
CFO Chat Agent	Orchestrates guardrail check → route → call agents → synthesize → validate, answering natural-language questions grounded in the other agents' findings	LangGraph state graph, Anthropic / Gemini (pluggable)

Guardrail-first, not bolted on

Guardrails — scope checking, prompt-injection input sanitization, and Pydantic output-schema validation — run before any agent or LLM call, so a blocked question never touches the database.
Multi-tenant isolation is enforced at three layers: every route derives organization_id from the verified JWT rather than client input, every query filters on org_id server-side, and calls to the agent service are rebuilt from the verified token, never from the client's raw request.
Every number is computed live from transaction history — no hardcoded financial data — and forecasts are checked against real held-out data rather than assumed accurate.

TalentOrch AI — Agentic Recruitment Platform

May — Jul 2026

Multi-agent platform automating the hiring lifecycle — 6 agents, LangGraph, Gemini API

Built a multi-agent recruitment platform that automates the hiring lifecycle — from resume screening and candidate evaluation to interview scheduling and hiring decisions — using a coordinated 6-agent LLM architecture.
Engineered the backend with LangGraph, FastAPI, Gemini API, FAISS, PostgreSQL, and MCP, implementing stateful agent orchestration, intelligent task routing, 5+ automated recruitment workflows, and human-in-the-loop escalation for critical decisions.
Replaced keyword search with a FAISS vector index for cosine-similarity matching across skills and experience, feeding an auditable, bias-aware decision process.

PythonFastAPILangGraph Gemini APIPostgreSQLFAISS MCPDocker

Pipeline — orchestrated by a LangGraph state machine

Job Posted → Resume Screening Agent (embed · FAISS index · bias-flag) → Orchestrator (score check — pass / auto-reject / escalate) → Interview Scheduling Agent → Pre-screening Agent (structured agentic chat) → Assessment Agent (auto-grades against a rubric) → Decision Support Agent (weighted recommendation + bias-risk check) → Communication Agent (stage-aware candidate messaging)

Agents — one file, one responsibility each

Agent	Responsibility	Key tech
ResumeScreeningAgent	Parses resumes, embeds & compares to the job description, flags PII so scoring stays bias-blind, indexes into FAISS	Sentence-Transformers, FAISS
InterviewSchedulingAgent	Finds mutual availability, sends invites, handles reschedules	Calendar service (mockable)
PrescreeningAgent	Runs a structured text-based pre-screen and extracts signals	Gemini API
AssessmentAgent	Administers and auto-grades technical assessments against a rubric	Gemini API
DecisionSupportAgent	Aggregates weighted scores into a recommendation, surfaces bias risk	Gemini API + rules
CommunicationAgent	Sends stage-aware, tone-controlled candidate messaging	Email service (mockable)

Built bias-aware from the start, not bolted on

PII, age, gender-coded terms, and graduation years are stripped before a resume ever reaches scoring, so the score reflects substance, not identity.
The Decision Support agent runs an aggregate parity check across flagged demographic proxies and surfaces a warning for human review if a group is screened out disproportionately — a signal, never an auto-decision.
Every agent decision is logged to a structured trace, so recommendations stay auditable after the fact.

github.com/sanskar024/talent-acquisition-agent

Advanced AI Chatbot with DeepEval & MCP

Feb — Apr 2026

Agentic RAG chatbot with hybrid retrieval, DeepEval-driven evaluation, and full observability

Engineered an agentic RAG chatbot with LangGraph, FAISS, LLMs, and MCP, supporting multi-document Q&A with source citations, live web search, stock-price retrieval, calculations, and autonomous tool orchestration with dynamic model routing for complex queries and LangSmith tracing for full observability.
Built a hybrid BM25 + FAISS retrieval pipeline with cross-encoder reranking and a self-correction loop for answer groundedness, using DeepEval to boost recall from 60% to 95% and precision from 50% to 85%.
Added human-in-the-loop stock validation, Redis-backed caching and rate limiting, and PostgreSQL-backed conversation memory for 50+ concurrent sessions.

PythonDeepEvalBM25 + FAISS LangChainPostgreSQLDocker RedisMCPLangSmith

Repo layout — sanskar024/Advanced_chatbot_langgraph

Advanced_chatbot_langgraph/ ├── chatbot_backend.py LangGraph agent logic, hybrid BM25+FAISS retrieval, tool routing, checkpointing ├── frontend_streamlit.py Streamlit chat UI, streaming responses ├── chatbot.py entry point └── chatbot.db / chat.db PostgreSQL/SQLite-backed thread & message persistence

github.com/sanskar024/Advanced_chatbot_langgraph

Model Regression Detection System

Nov 2025 — Jan 2026

CI/CD-style evaluation pipeline for LLM features

Built a CI/CD-style evaluation pipeline that continuously tests an LLM-powered email classification feature against a golden dataset and blocks deployment of quality regressions before they reach users.
Curated a 30-case hand-labeled golden dataset spanning 4 categories and deliberate edge cases (ambiguous intent, typos, mixed language, sarcasm), and built a multi-dimensional scorer combining exact-match accuracy with LLM-as-judge summary relevance.
Designed a diff engine comparing every new run against the last baseline, auto-flagging per-case regressions and improvements with configurable warning (3%) and critical (8%) thresholds.
Automated the pipeline into a GitHub Actions workflow that runs on every prompt-file PR, posts pass/fail status as a PR comment, generates an HTML diff report, and blocks merges on critical regressions; added rolling 7-run drift detection, Slack alerts, and Docker packaging.

PythonPydanticGitHub Actions DockerSlack Webhooks

github.com/sanskar024/-model-regression-detector

YoutubeRAG

Nov 2025 — Jan 2026

LLM-powered transcript analysis system

Designed an end-to-end RAG pipeline extracting YouTube transcripts and answering natural-language questions via FAISS-based semantic retrieval — processed 100+ videos at 85% answer relevance.
Built a chatbot interface surfacing context-aware answers grounded in transcript content, reducing average query response time to under 3 seconds.

PythonLangChainGemini API FAISSHugging FaceYouTube Transcript API

Repo layout — sanskar024/GenAi_youtube-chatbot

GenAi_youtube-chatbot/ ├── yotube_chatbot/ transcript extraction · chunking · FAISS retrieval · Q&A chain └── currency_conversion companion tool-use module (agentic function-calling demo)

github.com/sanskar024/GenAi_youtube-chatbot

Sanskar Gupta
builds systems that reason.

Where I've worked.

What I've been shipping.

Agents, pipelines, and tools.

The toolkit.

The record.

Sanskar Guptabuilds systems that reason.

Where I've worked.

What I've been shipping.

Agents, pipelines, and tools.

The toolkit.

The record.

Sanskar Gupta
builds systems that reason.