Bangalore, India

I build AI systems
that survive production.

Multi-agent pipelines. Production infra. End-to-end delivery.
30K → 200K users. Prototype → stock exchange.

AGENT_ARCHITECTURE / live
200K+
MAU scaled
30→97%
Accuracy gain
4
Live systems
~50%
Energy saved

> routing to capability_layer...

capability_layer /[6 nodes online]

What I Build

NODE_01
STATUS: ACTIVE

Multi-Agent Systems

Design and deploy production agent pipelines. Not wrappers. Full systems with memory, tools, routing, and evaluation.

utilization94%
NODE_02
STATUS: ACTIVE

Document Intelligence

High-recall RAG for complex unstructured docs. Built for compliance, finance, and enterprise domains where failure is not an option.

utilization97%
NODE_03
STATUS: ACTIVE

Voice AI & Real-Time

End-to-end voice pipelines — STT → LLM → TTS. Low-latency, stateful, handles real-world edge cases like interruptions and drift.

utilization88%
NODE_04
STATUS: ACTIVE

AI Infrastructure

From single server to Kubernetes. Canary deploys, Datadog monitoring, experimentation pipelines, production debugging.

utilization91%
NODE_05
STATUS: ACTIVE

Evaluation Systems

Eval frameworks that measure what actually matters. LLM-as-a-judge, funnel tracking, A/B pipelines for agent behavior.

utilization85%
NODE_06
STATUS: ACTIVE

0→1 Product Builds

Full-stack AI delivery. I own the surface: backend, infra, LLM layer, deployment. First user to 100K+ users.

utilization100%

> fetching production_deployments... 4 runs found. rendering results.

production_deployments /[4 runs found]

Run History

> fetching production_deployments... 4 runs found. rendering results.

STATUS: DEPLOYED
|run_id: drhp_bse_prod|2024-10|ENTERPRISE
Agentic RAGLangGraphBAMLHybrid RetrievalAWSEnterprise

DRHP Analysis System

First AI system deployed in production at BSE

One of India's first enterprise AI systems adopted inside a major stock exchange. End-to-end agentic RAG for DRHP compliance validation — built for high recall because in compliance, a miss is a failure.

EVAL_METRICS
parsing_success_rate
30% ──────────────── 97%
retrieval_strategy
dense + sparse hybrid
deployment
Kubernetes / AWS
users
BSE compliance teams
STATUS: LIVE
|run_id: oolka_mau_scale_prod|2025-12
Multi-AgentSpring BootPythonRedisDatadogGemini

Oolka AI — Multilingual Finance Chatbot

Rebuilt from scratch, scaled 30K → 200K+ MAU

Rebuilt Oolka's entire AI chatbot system from scratch — moved from a brittle single-agent setup to a scalable multi-agent multilingual architecture serving as a personal finance assistant for Indian consumers.

EVAL_METRICS
monthly_active_users
30K ──────────────── 200K+
architecture
multi-agent, multilingual
stack
Java orchestrator + Python LLM
observability
Datadog + A/B pipelines
STATUS: SHIPPED
|run_id: voice_ai_interview_freelance|2025-06
Voice AISTT/TTSReal-TimeLLM OrchestrationStreaming

Voice AI Interview Platform

Real-time AI interviewer with dynamic questioning and candidate eval

Built a real-time AI interview platform that conducts live voice interviews — not a static Q&A bot. The system dynamically generates follow-up questions, maintains conversation context, evaluates responses, and outputs structured assessments.

EVAL_METRICS
interaction_type
real-time voice (not async)
pipeline
STT → LLM → TTS loop
hard_problems_solved
latency, drift, edge speech
output
structured candidate evals
STATUS: DEPLOYED
|run_id: bert_labs_rl_energy_prod|2024-09
Reinforcement LearningSoft Actor-CriticAzure K8sMQTT

Industrial RL Energy Optimization

~50% energy cost reduction across HVAC, cement plants, airports

Built RL systems using Soft Actor-Critic for real-time industrial energy optimization across HVAC systems, cement plants, and airports. Deployed on Azure Kubernetes with MQTT for real-time sensor data streaming.

EVAL_METRICS
energy_cost_reduction
baseline ──────────── -50%
algorithm
Soft Actor-Critic (SAC)
deployment
Azure Kubernetes Service
data_pipeline
MQTT real-time streaming
200K+MAU scaled
30% → 97%Accuracy gain
~50%Energy reduction
deployedBSE production
4Live systems
pipeline_stages /[4 stages]

Experience

STAGE_04

ML Systems Engineer

Oolka AI
Dec 2025 – Present
INPUT:

broken single-agent chatbot, 30K users, no observability

TRANSFORM:

rebuilt multi-agent architecture from scratch, full-stack ownership, scaled infra, A/B eval pipelines, Datadog monitoring

OUTPUT:

200K+ MAU, scalable multi-agent system, canary deploys live

STAGE_03

Founding AI Engineer

OnFinance AI
Oct 2024 – Present
INPUT:

compliance team manually reviewing 500+ page DRHP documents

TRANSFORM:

designed and built agentic RAG system, BAML agent framework, hybrid retrieval, structured output enforcement, Kubernetes deployment on AWS

OUTPUT:

30% → 97% structured parsing success, deployed at BSE production

STAGE_02

Research Engineer

BERT Labs
May – Sept 2024
INPUT:

manual HVAC and energy controls in industrial environments

TRANSFORM:

built RL systems using Soft Actor-Critic, MQTT real-time data pipeline, Azure Kubernetes deployment across multiple facility types

OUTPUT:

~50% energy cost reduction across HVAC, cement plants, airports

STAGE_01

Research Intern

MITACS, Canada
May – Aug 2023
INPUT:

unstructured visual data requiring automated analysis pipeline

TRANSFORM:

designed computer vision pipeline, model training, evaluation framework, cross-domain generalization research

OUTPUT:

95% accuracy on benchmark, research pipeline deployed

tools_manifest /[available]

Tech Stack

REASONING_ENGINES
LangChainLangGraphLangSmithBAMLOpenPipeGeminiGPT-4Qwen
INFRASTRUCTURE
KubernetesDockerAWSAzureRedisDatadogGitHub Actions
RUNTIME
PythonJavaSpring BootFastAPIGunicornGevent
ML_PRIMITIVES
RAGRL (SAC, GRPO)Fine-tuningHybrid RetrievalVector DBsEvals
INTERFACES
Voice AISTT / TTSWebSocketsStreaming APIsREST
agent_properties /[3 properties]

How I Work

property: ownership

Tell me the outcome. I'll figure out the system.

I've gone zero-to-production on 3 live systems. I don't wait for a spec — I ask what you're trying to achieve and work backwards from there.

property: delivery_bias

I ship, then improve.

I've learned where to cut scope and where you absolutely cannot. Every project I've touched is live. Not in staging. Not in a demo. Live.

property: context_depth

I can talk to your CTO and your compliance team.

I understand architecture tradeoffs and I understand business requirements. I've navigated both in the same week — at a stock exchange.

> all_layers_traversed: true — generating output interface...

output_layer /[terminal]

Let's build something serious.

OUTPUT_TERMINAL — system: abhinavsingh2626@gmail.com
> connecting...

currently available for select projects  |  response_time: < 24 hours

abhinav.singh — AI Systems Engineer — Bangalore, Indiabuilt with systems thinking