Abhishek Sharma — I build AI systems that survive real users.
AI/ML Engineer specialising in agentic AI, production RAG and real-time voice. 2+ years shipping LLM systems across 50+ enterprise tenants.
Interactive 3D illustration of an agent network: agents, retrieval, voice and vision nodes around a central core.
On one multi-tenant RAG platform with zero cross-tenant leakage
Retrieval-to-answer latency across Arabic & English collections
Cut from 3 seconds to 200 ms with duplex speech-to-speech
Multi-camera re-identification across retail & QSR zones
Transitioned from experimental notebooks to live AWS microservices
Production AI that survives real scale.
Architected and deployed across enterprise tenants, low-latency WebRTC streams, and high-throughput computer vision pipelines.
Engineering for reality, not demos.
Systems designed with failure visibility, strict retrieval isolation, and deterministic performance guarantees.
“Most AI demos fall apart the moment real users touch them. I build the ones that don't. I work on agentic AI, retrieval and real-time voice systems that have to hold up in production — across tenants, languages and latency budgets. My focus: retrieval that returns the right context, not just similar-looking context; multi-agent workflows that don't silently break when one step fails; and systems that still work at 10,000 users, not just 10.”
Retrieval before generation
Get the right context in first; the model is only as good as what it's given.
Design for the failure path
Agents break at step 3 of 5. Build visibility, retries and fallbacks in.
Measure it
Latency, accuracy and answer quality get numbers and test sets, not vibes.
Engineering track record & measurable impact.
Proven track record taking multi-tenant RAG, conversational voice agents, and high-throughput computer vision models from architecture into enterprise production.
AI/ML Engineer (Agentic & Generative AI)
Active- Architected Synapse, a multi-tenant Arabic–English agentic RAG platform: LangGraph multi-agent orchestration, tenant-isolated Qdrant collections, retrieval-layer access control. 287 ms p95 latency across 50+ enterprise tenants; answer quality scored with DeepEval on a 90-question test set.
- Built a multilingual real-time speech-to-speech sales agent on LiveKit + Gemini Live (low-latency streaming, VAD, turn detection), autonomously handling 70% of client conversations.
- Replaced a cascaded STT→LLM→TTS interview pipeline with a duplex speech-to-speech agent (LiveKit + Gemini Live): turn latency 3 s → 200 ms, natural barge-in, extended to outbound SIP telephony — 25+ first-round screening calls/week.
- Rebuilt a multi-camera occupancy-analytics pipeline (YOLOv11 + BoT-SORT) for retail/QSR clients: fixed duplicate counts and identity loss by correcting ONNX ReID inputs, retuning the detection confidence threshold, moving inference to TensorRT FP16 with NVDEC decoding → 90% re-identification accuracy.
- Evaluated the real-time voice stack (LiveKit, Pipecat, Vapi, Retell) on cost, latency and Indian-language coverage; set org-wide defaults.
Python AI Developer
- Designed DECIMA, a GenAI + computer-vision pathology diagnostic system: detects/counts WBCs, RBCs, platelets from microscope images, classifies disease patterns, generates LLM-written PDF reports. 90% classification accuracy on 100+ labelled slides; review time 5 hours → 2 minutes.
- Shipped an NL2SQL ERP insights assistant (LangChain + LangGraph, schema-grounded query generation) so non-technical staff self-serve reporting, clearing a 2-day analyst request backlog.
- Containerised 15+ ML models as FastAPI microservices with Airflow-orchestrated retraining and CI/CD on AWS, moving all into production.
AI/ML Developer
- Built ML-driven interaction layers for an AR/VR metaverse platform using real-time data so virtual environments respond to user behaviour.
AI/ML Intern
- Built an end-to-end annotation pipeline turning architectural PDFs into labelled data; trained and fine-tuned YOLOv8 and Graph Neural Network models to predict building layouts from plans and geospatial data.
Artificial Intelligence Intern
- Fine-tuned a 'Mental Fitness Tracker' predictive model, improving wellness-outcome prediction accuracy by 21%.
Interactive technical capabilities & production usage.
Every capability backed by actual deployment context across multi-tenant architectures, real-time audio streams, and edge video pipelines.
Agentic AI & LLMs
Autonomous multi-agent orchestration, stateful graph architectures, and tool invocation pipelines.
Select or tap any skill above to view where and how it was deployed in production.
Agentic AI & LLMs
- LangGraph: Synapse multi-tenant RAG, NL2SQL ERP, & Legal AI workflows
- LangChain: Bilingual document retrieval chains and agent prompt pipelines
- Multi-Agent Systems: Orchestrated collaborative query decomposition and synthesis
- MCP (Model Context Protocol): Standardized agent tool invocation and context bridges
- Tool Calling: Function calling and schema enforcement across Gemini & GPT engines
- Prompt Engineering: Few-shot calibration, structured output parsing, and guardrailing
RAG & Retrieval
- Qdrant: Tenant-isolated vector collections serving 50+ enterprise tenants
- Pinecone: Serverless hybrid vector indexing for the Shriji Bhagavad Gita assistant
- pgvector: Relational + vector co-located query storage for enterprise records
- FAISS: In-memory similarity indexing for low-latency candidate clustering
- Hybrid Search: Sparse BM25 combined with dense semantic embeddings
- Jev / Reranking: Precision cross-encoder reranking driving 287 ms p95 Synapse queries
- Embeddings: Multilingual dense representations across Arabic, English, and Hindi
- LlamaIndex: Hierarchical document parsing and automated index chunking
Voice & Real-Time AI
- LiveKit: Low-latency WebRTC SFU streaming for duplex audio and video sessions
- Gemini Live: Native real-time speech-to-speech streaming with ultra-fast turnaround
- Pipecat: Evaluated for cost-sensitive multilingual conversational deployments
- Duplex Speech-to-Speech: Replaced cascaded STT-TTS, cutting latency from 3 s to 200 ms
- Silero VAD & Barge-In: Natural conversational interruption and real-time turn detection
- SIP Telephony: Outbound first-round screening agent running 25+ calls/week
- WebSockets & SSE: Streaming audio chunks and real-time sentiment observer metrics
Computer Vision
- YOLOv11: Real-time retail person detection deployed at 30 FPS on edge hardware
- BoT-SORT / ByteTrack: Multi-camera trajectory association achieving 90% re-ID accuracy
- TensorRT FP16: Edge GPU inference acceleration cutting latency by 2.8x
- ONNX Runtime: Cross-platform model exports and tuned ReID feature extraction
- OpenCV: Digital pathology microscope WSI tile preprocessing and stitching
- NVDEC Hardware Decode: Zero-CPU hardware video decoding for 8+ concurrent 1080p feeds
MLOps & Deployment
- FastAPI: Asynchronous high-throughput microservices for 15+ production models
- Docker: Hardened minimal base images for CPU and CUDA GPU workloads
- AWS (ECS/EC2/S3): Scalable cloud infrastructure hosting multi-tenant AI pipelines
- Apache Airflow: Automated DAGs for scheduled dataset ingestion and model retraining
- CI/CD & GitHub Actions: Automated linting, container builds, and zero-downtime deployment
- MLflow: Experiment tracking, model registry, and artifact versioning
- vLLM: High-throughput PagedAttention inference server optimization
- PEFT Fine-Tuning: Parameter-efficient LoRA tuning for domain-specific LLM adapters
Evaluation & Safety
- DeepEval: Automated test harnesses scoring faithfulness across 90-question test set
- RAGAS: Harmonized evaluation of context recall, precision, and answer relevancy
- LangSmith: Distributed trace inspection and agentic failure-point debugging
- LLM-as-a-Judge: Calibrated rubric-based quality auditing for clinical and sales outputs
- Safety Guardrails: Deterministic schema boundaries preventing hallucination in legal/clinical systems
- PII Detection: Automated redaction of sensitive patient and enterprise identifiers
Languages & Frameworks
- Python: Primary engineering language for AI backends, agents, and data pipelines
- TypeScript: Type-safe full-stack interface development and API contract consumption
- SQL & PostgreSQL: Relational modeling, complex joins, and pgvector operations
- PyTorch: Deep learning model development, fine-tuning, and inference pipelines
- Scikit-Learn: Predictive health outcome algorithms and classical ML classification
Spoken & Retrieval Languages
- English: Native-grade technical documentation and conversational AI
- Arabic: Synapse enterprise bilingual RAG across Middle Eastern legal & institutional corpora
- Hindi & Indian Languages: Multilingual voice agents & Kisan AI WhatsApp agriculture assistant
More selected projects & system builds.
Applied systems ranging from zero-cost serverless RAG and legal agent workflows to edge healthcare assistants and agricultural voice tools.
Health AI Assistant & Wellness Predictor
Healthcare AI assistant featured in Patrika newspaper; fine-tuned predictive wellness models improving outcome accuracy by 21%.
Legal AI Agent & Contract Assistant
Autonomous agentic workflow for contract clause auditing, risk analysis, and case precedent retrieval.
Kisan AI
WhatsApp-first multilingual AI assistant providing personalized crop diagnostics, weather alerts, and mandi prices for Indian farmers.
NL2SQL ERP Insights Assistant
Schema-grounded natural language SQL assistant enabling non-technical staff to self-serve ERP reporting.
MLOps & Microservice Productionisation
Productionisation of 15+ machine learning models into containerized FastAPI microservices with Airflow on AWS.
Real-time voice stack decision matrix.
Evaluated LiveKit, Pipecat, Vapi, and Retell across operating cost, latency budgets, and Indian-language dialect coverage to standardize production defaults.
Default for Browser-Embedded Real-Time & WebRTC Telephony
Default for Cost-Sensitive & Custom Multilingual Server-Side Agents
| Framework | Cost Profile | Latency Budget | Indian Language Coverage | Production Fit |
|---|---|---|---|---|
LiveKitProduction Default | Open-source / Self-hosted or Cloud | Ultra-Low (< 250 ms) | Broad (Model-dependent, Gemini Live native) | Browser-embedded WebRTC sessions, bidirectional video/audio, custom SFU infrastructure |
PipecatProduction Default | Highly Cost-Efficient (Open-source framework) | Low (~300–400 ms) | Excellent (Flexible STT/TTS routing for Indian languages) | Cost-sensitive multilingual deployments, telephony bots, and custom Python voice agent pipelines |
Vapi | Higher (Per-minute managed fee) | Low (~350–500 ms) | Good (Managed provider matrix) | Rapid enterprise prototyping and managed telephone workflows |
Retell AI | Higher (Per-minute managed fee) | Low (~400–600 ms) | Moderate (Standard global models) | Turnkey call center replacements and automated outbound voice campaigns |
Formal foundation & industry recognition.
Academic background in Information Technology, rigorous machine learning specializations, and regional media recognition.
Bachelor of Technology (B.Tech)
Rajiv Gandhi Proudyogiki Vishwavidyalaya (RGPV) · Bhopal, India
- Machine Learning SpecializationTopStanford Online (Coursera)
- Machine Learning & AI with Hands-on ProjectsTopIndustry Accredited
- Machine Learning & Deep Learning in Python & RAdvanced AI Specialization
- Introduction to Programming Using PythonSoftware Foundations
- Data VisualisationAnalytics Program
- AWS Summit India ParticipationAmazon Web Services
Featured for Healthcare AI Assistant
Recognized in regional media for building an accessible healthcare AI diagnostic assistant providing early wellness insights.
Start a conversation & hire Abhishek.
Available for AI Engineer, Applied AI Engineer and Agentic/GenAI roles (Remote or India). Let's build production systems that endure.
Direct inbox access
Prefer writing directly from your own email client or connecting on professional networks?