Ajay Jangid
AI & GenAI Engineering · Mumbai

Ajay Jangid

I architect production AI for banking — voice analytics, retrieval and agents that have to hold 99.9% uptime, run on self-hosted GPUs inside an air-gapped network, and stand up to an auditor.

AVP, AI & GenAI Engineering at Axis Bank 7+ years shipping ML to production M.Tech ML & AI, BITS Pilani
See what I'm building Résumé (PDF) LinkedIn
1.2M+customer calls analysed every month
18,000+contact-centre agents on the platform
60K+AI inference requests per day
99.9%production SLA, zero regulatory observations
01 · About

From models to platforms

I build the systems that sit between a model and a regulated business. At Axis Bank I lead AI & GenAI engineering — the platforms that turn millions of customer conversations, documents and workflows into decisions, with the governance a bank needs: PII masking, role-based access, audit trails, and controls aligned to KYC/AML.

My work spans the whole stack: self-hosted speech recognition for noisy Indic telephony audio, vLLM serving tuned for throughput on a fixed GPU budget, retrieval with guardrails, agentic workflows, and the evaluation harnesses that tell us whether any of it is actually getting better. I also write the business case — GPU economics and cost-per-call — because architecture decisions in a bank are budget decisions too.

Before Axis I was at Accenture AI, BRIDGEi2i and a run of product teams, building recommendation engines, insurance-claims extraction, and computer vision. I've been writing on Medium about applied ML since 2019, and I mentor data scientists and interns on the team.

02 · Now

What I'm working on

  • INFERENCE

    Throughput, not latency

    Pushing calls-per-day capacity on a fixed 4×A10G budget — independent vLLM replicas, prefix-cache reuse, length-bucketed batching and KV-cache offload experiments.

  • SPEECH

    Indic telephony ASR

    Benchmarking and fine-tuning ASR for 8 kHz, code-switched call audio: IndicConformer, Whisper large-v3 and ARTPARK-IISc's SraVaani, plus joint ASR + diarization models.

  • AGENTS

    Agents for the next earnings call

    Taking the IR Question Intelligence Platform across more banks — forecasting analyst questions, with held-out-quarter evals to prove it generalizes.

  • AGENTS

    Agentic workflows for banking ops

    Text-to-SQL, email triage and risk & compliance agents — with the harness engineering, observability and evals that make agents safe to ship.

  • TEACHING

    Architect-level curricula

    Expanding two open, interactive guides on agentic AI and RAG architecture, written from what actually breaks in production.

Last updated September 2026
03 · Systems

Selected work

Flagship · Axis Bank · 2023 – present

Call Analytics Intelligence Platform

An enterprise platform that listens to every contact-centre call and turns it into intent, sentiment, compliance flags, a 20-parameter quality score and agent-level performance trends. Around 2,000 call-hours a day, on a model-agnostic stack.

REC Namaste! How can I help? Mera card block ho gaya… Unblocking it right now. Thank you so much! The callRM ↔ customer, recorded Telecom networkPSTN · SIP trunk Bank networkfirewall · recording server +1 Secure storageencrypted recordings AI framework ASR LLM Score model-agnostic · PII masked AI frameworktranscribe · understand · score 17/20 ⚑ disclosure missed Business dashboardscores · flags · trends 17/20 ⚑ Noted: explain charges next time. RM gets feedbackcoaching · WoW trends ♪ rec
  1. The call An RM helps a customer; the conversation is recorded.
  2. Telecom network The audio travels over PSTN / SIP trunks into the bank.
  3. Bank network Inside the firewall, the recording server captures the call.
  4. Secure storage Recordings land in encrypted storage, ready for analysis.
  5. AI framework Pluggable ASR transcribes, an LLM scores the call on a 20-point rubric, with PII masked throughout.
  6. Business dashboard Scores, compliance flags and trends flow to the business.
  7. Feedback loop The RM sees exactly what to improve on the next call.

Architecture

Model-agnostic by design: a config-driven router puts open-source models (served on vLLM) and managed APIs (OpenAI, Gemini, Bedrock) behind one interface with fallback chains — for ASR and LLMs alike.

Quality

A Cohen's κ inter-rater harness measures model scoring against human auditors on every rubric parameter before changes ship.

Impact

1.2M+ interactions a month for 18,000+ agents; relationship-manager analytics across ~60K calls/day and 10K+ RMs.

Agentic AI · Investor Relations · 2026

IR Question Intelligence Platform

AI agents for the next earnings call. The platform predicts which questions analysts will ask, profiles each analyst's recurring themes across quarters and banks, and grounds every answer in transcripts and disclosures — then turns it all into a call-ready brief. It started as a knowledge-graph prototype; when it overfit to a single quarter, I rebuilt it as an agentic system with the graph as long-term memory and an eval harness to test generalization.

Outlook on credit costs? We see them normalising. Earnings callanalyst ↔ management Sourcestranscripts · decks · filings credit costs Analyst A open thread Knowledge graph memoryanalysts · topics · banks · threads Agent harness F Forecast A Analyst E Evidence B Brief 4 agents · shared memory Agent harnessforecast · profile · ground · brief Call-prep brief #1 Credit costs outlook ↳ Analyst A follow-up ↳ 2 citations attached ↳ answer drafted Call-prep briefranked questions · citations Credit costs next quarter? Here's our data on that. ✓ predicted #1 Next earnings callthe team is ready rec
  1. Earnings call Analysts question management, and every call becomes a transcript.
  2. Sources Transcripts, investor decks and filings are ingested every quarter.
  3. Knowledge graph memory Analysts, topics, banks and open threads are linked across quarters.
  4. Agent harness Forecast, Analyst, Evidence and Brief agents read from memory, reason, and write back.
  5. Call-prep brief Ranked questions, analyst context and citations, ready for the IR team.
  6. Next earnings call The predicted question comes up and the team is ready. Held-out quarters measure how often that happens.

Evaluation

Held-out quarters as the test set, question-recall scored against what analysts actually asked, and a backtesting loop for every weight change.

Operations

Continual learning with logged proposals, and an audit log of every LLM call across providers.

Coverage

Live workspaces for Axis Bank, Kotak Mahindra and IndusInd, with a repeatable pipeline to add more banks.

LangGraphFastAPIHybrid searchKnowledge graphOpenAI · Gemini · Groq
Open the live demo →View code on GitHubRead the write-up (2 parts)
Retrieval

Enterprise RAG platform

Production retrieval-augmented generation over customer calls and documents — dense + hybrid search, reranking, knowledge graphs, and guardrails with hallucination control and eval/safety gates.

80%+ extraction accuracy · ₹2 Cr+ annual savings
GPT-4GeminiLlamaPineconeWeaviateKendra
Agents

Agentic banking workflows

Multi-agent automation for core operations: text-to-SQL with schema grounding and query validation, email triage and drafting, and risk & compliance agents for regulatory mapping and audit trails.

Analytics, ops and compliance workflows automated
CrewAILangChainReActPlan-and-Execute
Extraction

Entity & information extraction engine

Fine-tuned, LoRA-adapted BERT models replaced an incumbent vendor for structured extraction from unstructured banking text.

85%+ F1 · cost cut from ₹1.2 Cr to ~₹30 L a year
TransformersLoRAPEFT
Marketing

Customer intelligence & personalization

In-house GenAI platform for campaign generation and personalization, bringing work previously done by external vendors inside the bank.

+15% campaign engagement · ₹50 L+ saved a year
GPT-4GeminiLLMOps
04 · Client work

Freelance projects

Alongside my role at the bank, I build real-time voice and video agents for clients — the same production discipline, applied to conversations that happen live. Both projects below were built end to end on Pipecat.

Freelance · Healthcare · Voice AI

AI voice receptionist for clinics

A phone agent that answers a clinic's calls around the clock. It books, reschedules and cancels appointments, answers questions about timings, fees and location, sends confirmations and reminders, and hands the call to staff when a human is needed. Built on Pipecat as a fully streaming pipeline, so the conversation feels natural — even when patients interrupt.

✓ Confirmed · Thu 11:30 Pipecat voice pipeline DIALOG STATE intent: — specialty: — slot: — barge-in · TTS stopped VAD STT LLM TTS streaming · interruptible · tool-calling Clinic calendar Thu 11:30 Fri 16:00 Patient recordsbooking saved ✓ SMS / WhatsAppconfirmation sent ✓ Patientcalls the clinic TelephonySIP · WebSocket Voice agentVAD → STT → LLM → TTS Toolscalendar · records · SMS LIVE TRANSCRIPT Patient Hi, I need to see a cardiologist this week. ⚙ check_availability(cardiology, this_week) → Thu 11:30 · Fri 16:00 Agent Our cardiologist is free Thursday 11:30 or Friday 4 pm. Which suits you? Patient Thursday, please! ↯ interrupted the agent ⚙ book_slot(Thu 11:30) ✓ ⚙ send_sms(confirmation) ✓ Agent Done! You'll get an SMS confirmation in a moment.
  1. Patient calls Voice activity detection and streaming speech-to-text pick up the request as it's spoken.
  2. LLM + tools The agent understands the intent and checks the clinic calendar for open slots.
  3. Agent replies Text-to-speech streams the answer back in a natural voice.
  4. Barge-in The patient interrupts; the agent stops talking instantly and listens, like a good receptionist.
  5. Booking The slot is saved in patient records and an SMS confirmation goes out.
  6. Done One short call, no hold music, available 24/7 — with a human handoff when needed.

Real-time pipeline

Streaming VAD → STT → LLM → TTS in Pipecat, with barge-in handling and latency tuned stage by stage across speech and LLM providers.

Tool use

Function calls into the clinic calendar, patient records and SMS/WhatsApp, with the agent confirming details before anything is booked.

Beyond booking

Reschedules, cancellations, reminders, FAQs on timings, fees and location, and a clean handoff to clinic staff.

PipecatStreaming STTLLM function callingTTSSIP / WebSocketPython
Freelance · HR tech · Multimodal agents

AI interviewer: real-time voice + video

A multi-agent system that runs live job interviews. An AI interviewer talks with the candidate over video, adapts each question to their resume and answers, and a panel of LLM judges scores every response in parallel. Recruiters get a structured report with the full recording and transcript. It picks up a problem I first worked on in 2019 at Analysed.in — rebuilt for the LLM era.

RTC INTERVIEW · LIVE REC AI interviewer Describe a RAG system you built. you Pipecat multi-agent P Planner I Interviewer ×3 Judges R Recorder 4 agent roles · one conversation Candidatejoins from the browser WebRTClive audio + video Interview roomAI avatar · live captions Agent teamplan · ask · judge · record INTERVIEW LOG Planner resume + job description → Q1: Describe a RAG system you built. Interviewer asks Q1 · voice + avatar video + live captions Candidate “Hybrid retrieval — BM25 plus dense vectors, then a reranker…” Judges ×3 depth 4/5 · clarity 5/5 · relevance 4/5 · scored in parallel Planner follow-up → How did you evaluate retrieval quality? Interview report Technical depth Communication Problem solving ✓ Recommend: next round ▶ recording · transcript · per-question scores
  1. Candidate joins A browser call over WebRTC: the AI interviewer appears on video and recording starts.
  2. Question planning A planner agent reads the resume and job description and picks the next question.
  3. AI asks The interviewer agent speaks through a synced avatar, with live captions.
  4. Candidate answers Streaming speech-to-text follows the answer in real time, with natural turn-taking.
  5. Parallel judges Several LLM judges score the answer at once, and the planner adapts with a follow-up.
  6. Report Recruiters get competency scores, a recommendation, and the full recording and transcript.

Multi-agent design

Planner, interviewer, judge and recorder roles coordinated in Pipecat over WebRTC, with questions generated from the resume, the job description and earlier answers.

Evaluation

LLM-as-judge scoring against a per-competency rubric, with judge calls run in parallel so feedback is fast and more consistent.

Engineering

Solved WebRTC connectivity, race conditions and turn-taking latency; audio and video recording pipelines; deployed on AWS EC2 with a FastAPI backend.

PipecatWebRTCFastAPILLM-as-judgeAvatar videoAWS EC2
05 · Approach

How I build

Measure first

No model change ships without an eval that agrees with human reviewers. Better means faster inference, higher recall, lower hallucination rate — in numbers.

Own the inference

In a bank, data residency decides the architecture. Self-hosted serving, tuned for the GPUs you actually have, beats a slide about the GPUs you wish you had.

Compliance is a design input

PII masking, access control and audit trails are designed in from day one — which is how you reach production with zero regulatory observations.

06 · Writing

Guides & writing

23chapters

Agentic AI Architect

An interactive learning portal on designing and architecting production agentic systems — patterns, orchestration, harness engineering, observability and evals.

Read the portal →
27lessons

The RAG Architect's Handbook

Retrieval-augmented generation from first principles to production architecture: chunking, hybrid search, reranking, evaluation and scale.

Read the handbook →

Essays on Medium

All stories →
Aug 2026Agent Governance & Control: The Real Bottleneck in Agentic AI's Next PhaseAgentic AI Aug 2026Sovereign AI: Why “Where Your Model Runs” Became a Boardroom QuestionAI infrastructure Aug 2026Under the Hood: What Actually Made an AI-Powered IR System Work (Part 2)Knowledge graphs Aug 2026Beyond the Earnings Call: How AI Can Turn Investor Relations Into a Strategic Radar System (Part 1)Investor relations Aug 2026Razorpay Vulcan: What It Actually Means to Build a “Foundation Model” for PaymentsPayments AI Jun 2023Revolutionizing Investment Research: Unleashing GenAI Power Through Financial Data and NewsGenAI Feb 2023Transforming Unstructured Text into Structured Insights with NLP TechniquesNLP Sep 2019Intuitive Understanding of Seq2seq Models & Attention Mechanism in Deep LearningDeep learning · Analytics Vidhya Aug 2019Pfam Protein Sequence Multi-class Classification Using Deep LearningDeep learning
07 · Experience

Where I've worked

2023 – now

Axis Bank · Mumbai

AVP, AI & GenAI Engineering (Apr 2026 –) · Senior Data Science Manager (2023 – 26)

Lead the bank's AI engineering and GenAI platform: voice AI, RAG, agentic workflows, self-hosted LLM/ASR infrastructure and the MLOps/LLMOps lifecycle. Lead and mentor a team of data scientists and ML engineers.

2022 – 2023

Accenture AI · Senior Data Scientist

Graph-based anomaly detection (+25% accuracy, −30% false positives) and transformer-based search and ranking for emerging risks in insurance claims.

2021 – 2022

BRIDGEi2i Analytics · Data Science Consultant

Claims extraction engine that cut processing from 1–2 days to minutes; OCR + NLP over 10K+ documents a month at 98% entity accuracy.

2020

Code Matrix Software · ML Engineer

Hybrid semantic + collaborative-filtering recommender on Elasticsearch at 10M+ queries/day, sub-200 ms, 99.9% uptime.

2019 – 2020

Analysed.in · ML Engineer

An AI-powered virtual interview platform: resume parsing, skill extraction, candidate–job matching and scoring.

2018 – 2019

STYLABS · ML Engineer

GANs for image synthesis and YOLO-based object detection for real-time video analytics.

Education

  • M.Tech, Machine Learning & AIBITS Pilani · 2022 – 2024
  • B.Tech, Electronics & TelecommunicationShree L.R. Tiwari College of Engineering · 2014 – 2018

Recognition

  • First Runner-Up, Accenture Data Science Hackathon50+ teams nationally
  • Spot Award, AccentureHigh-impact AI delivery
  • Quarter Team Award, BRIDGEi2iExcellence in AI innovation
08 · Toolkit

What I work with

GenAI & agents

RAG (dense + hybrid), SFT / LoRA / RLHF, CrewAI, LangChain, LlamaIndex, text-to-SQL, tool-use pipelines

Inference

vLLM, TGI, Triton, KV / prefix caching, GPTQ · AWQ · GGUF quantization, multi-GPU replicas

Speech

Whisper fine-tuning, CTranslate2, IndicConformer, streaming ASR, diarization, telephony audio

Retrieval

FAISS, Pinecone, Weaviate, Chroma, AWS Kendra, Elasticsearch, cross-encoder reranking, knowledge graphs

MLOps & LLMOps

SageMaker, Azure ML, Cloudera CML, MLflow, Docker, Kubernetes, prompt versioning, LLM tracing

Data & code

Python, SQL, PySpark, Kafka, PostgreSQL, MongoDB, FastAPI, PyTorch, Hugging Face

Early projects · 2018 – 2019
Machine learning & NLP Deep learning Computer vision All code on GitHub →