Available · Remote · AI Systems Engineering

Raj Jaiswal

I build agentic AI systems that know when to stop and ask a human.

Production-grade agents with durable checkpointing, CI-gated hallucination detection, and full observability — not notebook demos.

Production agents. Durable. Observable. Not notebook demos.

IIT Roorkee 5★ Upwork · Verified Founder · Synlitics 3 systems in production Rust OS contributor

View Systems terminalGitHub

Production Systems

Three agentic AI systems deployed.

LangGraph · PydanticAI · FastMCP

Client Work

3 systems shipped for paying clients.

AI agents · RAG · Workflow automation

$ stack --current

LangGraph + PydanticAI + FastMCP
pgvector + Logfire + Docker

$ status

Available · Remote

Production Systems

Three agentic AI systems deployed.

LangGraph · PydanticAI · FastMCP

PRODUCTION SYSTEMS

Systems I built. Decisions they prevented.

Each system is production-grade, containerized, and observable end-to-end.

swipe to explore →

Hallucination Rate

<5%

CI-gated · 30 labelled contracts

Cost per Request

$0.011

↓ 74% via semantic caching

Agent Uptime

99.7%

Durable checkpointing

CVEs Caught

Transitive dependency scan

Client Systems

Shipped · Running in production

Production system Legal-risk AI agent

Contract Intelligence Engine

LangGraph PydanticAI pgvector Docker

What it is

An AI agent that reads contracts, scores risk, and stops itself before making a decision it isn't confident in.

What I built

Structured extraction pipeline using PydanticAI, pgvector hybrid retrieval over 500+ historical clauses, and a risk scorer that triggers interrupt_before at score > 40.

The hard part

Deciding when the agent should stop trusting itself. interrupt_before on a risk score above 40 was the line between a useful tool and a liability.

Why it works

Hallucination rate under 5% — CI blocks any merge that exceeds it. AsyncPostgresSaver means a mid-execution crash loses nothing.

Production system Autonomous CVE auditor

Dependency Drift Sentinel

LangGraph pgvector GitHub Actions

What it is

A GitHub Actions agent that catches vulnerable dependencies before they merge — and keeps re-checking until the vulnerability is actually gone.

What I built

Cyclic LangGraph loop resolving the full transitive dependency tree, cross-referenced against OSV.dev plus a semantic CVE search layer via pgvector.

The hard part

OSV.dev alone misses aliased package names. Layering semantic search over the CVE corpus is what caught the cases the deterministic check didn't.

Why it works

47 CVEs caught across scans. Recursive self-validation: the agent re-checks its own migration suggestions, not just the original tree.

Production system Zero-risk AI rollout

Shadow-Mode AI Deployment Proxy

FastAPI asyncio LLM-as-judge

What it is

Infrastructure that lets a team run a new AI agent against 100% of real production traffic with zero risk to live users.

What I built

asyncio.gather fan-out to legacy + shadow in parallel, three-layer divergence detection (schema → embedding similarity → LLM-as-judge), nightly k-means clustering of failures.

The hard part

Cutting per-request cost from $0.043 to $0.011 via semantic caching — not optimization for its own sake, but what makes continuous shadow testing affordable rather than a pre-launch ritual.

Why it works

Zero latency added to real users. CI gate requires ≥95% agreement before promoting the shadow agent to production.

CLIENT SYSTEMS

Shipped for real clients. Running in production.

swipe to explore →

Client delivery Workflow automation

AI Decision Layer with Slack Integration

LangGraph FastAPI Slack API PydanticAI

THE SITUATION

The client's team was manually reviewing every incoming decision — even the obvious ones. No triage, no prioritisation. High-confidence cases and edge cases landed in the same queue and got the same human time.

WHAT I BUILT

A LangGraph routing layer that classifies each decision by confidence, auto-approves above the threshold, and sends only the uncertain cases to a human via Slack. Full audit trail on every decision, approved or automated.

WHAT THEY GOT

Manual review volume cut significantly — humans now only see edge cases, not rubber stamps. Response time on escalations dropped because the Slack alert arrives with full context already attached, not a raw request.

Client delivery Document intelligence

Document QA with Citations & Analytics

pgvector FastAPI PydanticAI

THE SITUATION

The client's team was wasting hours searching through dense PDFs and word documents for specific clauses. When they tested off-the-shelf AI tools, they hallucinated answers and couldn't point to the exact source paragraph, making them untrustworthy.

WHAT I BUILT

A hybrid-search RAG pipeline using pgvector that strictly grounds answers in the source material. It generates precise, paragraph-level citations for every claim and automatically logs queries that the knowledge base couldn't answer.

WHAT THEY GOT

Immediate, verifiable answers to complex queries without reading full documents. The automated "Content Gap" report surfaced critical blind spots in their documentation that they didn't even know existed.

Client delivery Sales automation

Lead Classifier & CRM Automation

LangGraph PydanticAI CRM API

THE SITUATION

Sales reps were drowning in inbound leads, spending hours manually qualifying them. High-value prospects were sitting in the same inbox as spam, causing slow response times and lost revenue opportunities.

WHAT I BUILT

A multi-factor scoring engine that evaluates firmographic signals and intent. High-value leads are instantly enriched and pushed to the CRM. Edge cases trigger smart escalation rules for human review, reducing false negatives to near-zero.

WHAT THEY GOT

Sales team now only talks to pre-qualified, highly-enriched leads. Manual triage time dropped to near-zero, and high-value prospects receive immediate outreach, directly increasing close rates.

REAL WORK

Experience

Building AI systems and shipping production code.

domain

Founder & AI Systems Engineer

Synlitics · 2024–Present

› Architected production RAG pipelines for contract analysis
› Built autonomous CVE auditing agent with recursive self-validation
› Designed shadow-mode deployment infrastructure reducing AI rollout risk to zero
› Implemented durable checkpointing and human-in-the-loop patterns

"Excellent technical expertise and strong problem-solving skills. His ability to turn complex data into clear, actionable insights was impressive."

Suraj K. · Kivi Labs ✔ Verified

work

AI Systems Engineer — Freelance

Independent · 2024–Present

› Delivered production AI agents for workflow automation, document intelligence, and lead qualification
› Consistent 5-star client feedback across all engagements
› Built three client-facing systems currently running in production: decision routing layer, document chatbot with citation tracking, lead classifier with CRM automation

"Raj is professional, communicates clearly, and understands requirements well. Delivered quality work on time."

Asif Ali K. · Q4 Sales Analysis ✔ Verified

APPROACH

How I build

psychology

"An agent is only as good as its ability to know what it doesn't know."

I build systems that are traceable, interruptible, and honest about uncertainty.

gavel

An agent that can't say "I'm not sure" isn't production-ready.

ERR: CI gates block merges above 5% hallucination rate

route

Every agent decision should be completely traceable.

TRC: Logfire + OpenTelemetry distributed tracing active

monitoring

Cheap and observable always beats clever and opaque.

OPT: 74% cost reduction achieved via semantic caching

gavel

An agent that can't say "I'm not sure" isn't production-ready.

ERR: CI gates block merges above 5% hallucination rate

route

Every agent decision should be completely traceable.

TRC: Logfire + OpenTelemetry distributed tracing active

monitoring

Cheap and observable always beats clever and opaque.

OPT: 74% cost reduction achieved via semantic caching

CAPABILITIES

Technical stack

Grouped by system layer.

swipe to explore →

memory

Agentic Systems

LangGraph PydanticAI FastMCP interrupt_before AsyncPostgresSaver human-in-the-loop

RAG & Retrieval

pgvector hybrid search semantic caching document chunking context window mgmt

visibility

Observability & Evals

Logfire OpenTelemetry pydantic-evals LLM-as-judge hallucination CI gates

dns

Infrastructure

FastAPI Docker Redis PostgreSQL GitHub Actions

memory

Agentic Systems

LangGraph PydanticAI FastMCP interrupt_before AsyncPostgresSaver human-in-the-loop

RAG & Retrieval

pgvector hybrid search semantic caching document chunking context window mgmt

visibility

Observability & Evals

Logfire OpenTelemetry pydantic-evals LLM-as-judge hallucination CI gates structured logging

dns

Infrastructure

FastAPI Docker Redis PostgreSQL asyncio GitHub Actions CI/CD

WHAT OTHERS SAY

Feedback from freelance and consulting work

format_quote

"Professional, responsive, and delivered high-quality work on time. Communication was clear and attention to detail was excellent."

Shubham G. OECD Data Analysis Project

Verified

format_quote

"Reliable, communicates well, and completes tasks efficiently. I'd be happy to work with Raj again."

Client Review Project Engagement · Upwork

Verified

EARLIER WORK

Earlier systems

Analytics and ML work — retained for range.

swipe to explore →

Churn Prediction System

AUC 0.71 → 0.84 · XGBoost, SMOTE

50K users arrow_outward

Sales Forecasting

MAPE 14.1% → 8.3% · ARIMA, Prophet

Retail data arrow_outward

Restaurant Analytics SaaS

Automated ELT pipeline · Supabase, Power BI

Daily KPIs arrow_outward

A/B Testing Framework

+18% conversion lift · Hypothesis testing

Checkout flow arrow_outward

Churn Prediction System

AUC 0.71 → 0.84 · XGBoost, SMOTE

50K users arrow_outward

Sales Forecasting

MAPE 14.1% → 8.3% · ARIMA, Prophet

Retail data arrow_outward

Restaurant Analytics SaaS

Automated ELT pipeline · Supabase, Power BI

Daily KPIs arrow_outward

A/B Testing Framework

+18% conversion lift · Hypothesis testing

Checkout flow arrow_outward

terminalLET'S TALK

Let's build something.

Available remotely

Production-minded. Systems-oriented. Available remotely.

Get in Touch send GitHub code LinkedIn link