
Software Engineer – AI/ML Platform (AI/ML Engineering Lead)
Carvewing Solutions LLP · Jan 2026 – Present
- Led execution across backend, AI, CV, and enterprise automation products while coordinating a 5-engineer squad across HRMS, CAD automation, CLM/SAM, financial systems, and visual analytics platforms.
- Recognized as Employee of the Month within the first 3 months for technical execution, ownership, and delivery velocity.
- Co-led an AI-Assisted Project Planning workshop for 100+ attendees across internal teams, clients, partners, and engineering professionals.
- Designed and reviewed production AI workflows spanning multimodal extraction, agentic tool use, evaluation checkpoints, human feedback, safety controls, and measurable reliability/cost trade-offs.
- Defined production GenAI integration patterns for client solutions: secure identity/access boundaries, auditability, data sovereignty, guardrails, versioned prompt/configuration releases, rollback paths, and robust provider-failure handling.
- Designed deployment and retrieval architectures spanning AWS Bedrock, ECS/EKS, OpenSearch/vector-search patterns, Docker/Kubernetes, and self-hosted vLLM/Ollama—balancing latency, throughput, token efficiency, cost, and enterprise data boundaries.
- Translated client requirements into measurable LLM adaptation and evaluation plans, including the appropriate use of CPT, SFT, LORA, QLORA, and PEFT; actively implementing the corresponding fine-tuning experimentation workflow.
- Instrumented production services with structured logs and Prometheus/Grafana dashboards for p50/p95/p99 latency, error rate, throughput, queue depth, and background-job reliability.
- GenAI / Agents: LLM system design, agent orchestration, LangGraph, LangChain, AutoGen, Pydantic AI, tool calling, prompt and context engineering, Mem0 + Redis memory patterns, guardrails, prompt-injection mitigation, human-in-the-loop systems, RAG, embeddings, vector search, retrieval optimization, LLM evaluation and benchmarking, OpenAI APIs, Azure OpenAI, Groq, NVIDIA NIM, Ollama, Qwen, Whisper
- Quality / Observability: Langfuse, structured evaluation and benchmarking, human evaluation, retrieval failure analysis, error analysis, A/B testing, inference cost/latency/throughput optimization, token efficiency, structured logging, metrics, dashboards, OpenTelemetry concepts, production monitoring, robust failure handling
- Built an end-to-end engineering drawing automation platform in FastAPI and React/TypeScript that ingests TIF/PDF files, extracts annotations/symbols/dimensions, detects redline corrections, and presents proposed changes for interactive review and approval.
- Combined OpenCV preprocessing with EasyOCR/PaddleOCR, structural layout parsing, drawing-element identification, and Ollama/Groq-assisted token classification to generate structured correction metadata.
- Built repeatable document-evaluation flows around OCR quality, spatial consistency, token classification, and reviewer feedback; used failure analysis to improve robustness on noisy scanned engineering drawings.
- Traceability: Designed a manifest-first workflow that keeps source drawings immutable and records derivative outputs, annotations, approval state, and audit metadata across the review lifecycle.
- Safety: Enforced human-in-the-loop boundaries so AI performs extraction and recommendation while final engineering correctness remains explicitly user-approved.
- Built deterministic CAD-update and engineering-drawing generation workflows with validation, traceability, and human-review checkpoints.
- Architecture: Designed hybrid cloud/on-prem deployment patterns for customers with strict data-sovereignty and IP-protection requirements, evaluating self-hosted model-serving options for secure inference.
- Challenges solved: Improved robustness on noisy scanned drawings, preserved spatial relationships during extraction, converted irregular tabular engineering inputs into valid vector outputs, and supported secure local validation workflows.
- Tech: FastAPI, Python, React, TypeScript, OpenCV, EasyOCR, PaddleOCR, PyMuPDF, Ollama, Groq, OpenPyXL, XlsxWriter, SolidWorks API.
- Built and launched an enterprise HRMS with React/TypeScript and FastAPI covering employee management, leave, timesheets, multi-manager approvals, attendance, notifications, audit flows, KRA, skills, helpdesk, and administration.
- Integrated an LLM tool-calling layer that converts natural-language requests into controlled actions across leave, employee queries, timesheet status, and administrative workflows.
- Implemented constrained action schemas, validation boundaries, audit logs, and authorization checks around LLM-triggered actions to make enterprise agent workflows safe and reproducible.
- Migration: Orchestrated a zero-downtime SQL Server/MySQL-to-PostgreSQL migration across multiple enterprise data domains, with validation, rollback safety, and million-row-scale workloads.
- Security & reliability: Implemented RBAC, stateless JWT authentication, rate limiting, validation boundaries, auditability, scheduled automations, and production error alerting.
- Operations: Added service and job observability covering p95/p99 latency, failed/retried jobs, scheduled automation health, and production error alerts for post-go-live support.
- Challenges solved: Reconciled inconsistent legacy schemas and orphaned records, preserved approval relationships, reset database sequences safely, and stabilized post-go-live workflows without service interruption.
- Tech: React, TypeScript, shadcn/ui, FastAPI, PostgreSQL, SQL Server, MySQL, JWT, Docker.
- Built an end-to-end diagram-to-EDA reconstruction system that converts PDFs and scanned schematics into structured components, labels, tokens, connectivity mappings, and machine-readable circuit representations.
- Combined OpenCV preprocessing, EasyOCR, PyMuPDF rendering, and Ollama/Qwen-assisted classification behind FastAPI services, with a ReactFlow editor for correction and validation.
- Correctness: Designed an immutable, netlist-first architecture with human validation checkpoints before any reconstructed circuit is accepted for downstream EDA use.
- Challenges solved: Recovered multi-point connectivity from noisy raster inputs, reconciled OCR tokens with spatial geometry, and kept visual edits synchronized with structured netlist state.
- Tech: FastAPI, React, TypeScript, ReactFlow, EasyOCR, OpenCV, PyMuPDF, Ollama, Qwen 2.7B/7B, Python.
- Built an enterprise CLM/SAM platform in FastAPI/PostgreSQL covering contract creation, versioning, metadata and clause extraction, configurable state transitions, multi-party approvals, and immutable audit history.
- Integrated Groq-powered natural-language actions and asynchronous Slack/Teams webhooks so users can create, update, clarify, approve, or reject contracts without leaving chat.
- Workflow integrity: Enforced explicit lifecycle states, approval authorization, immutable event history, and strict API boundaries around every LLM-triggered action.
- Interoperability: Designed migration-ready schemas aligned with Icertis, Ironclad, SAP Ariba, and Conga, plus MuleSoft/Informatica-style NLP transformation pipelines.
- Challenges solved: Preserved auditability across conversational actions, normalized heterogeneous contract formats into one schema, and validated asynchronous integrations through secure local test environments.
- Tech: FastAPI, PostgreSQL, SQLAlchemy, Groq LLM, Slack API, Microsoft Teams API, REST APIs, JSON workflows, Docker.
- Built a live CCTV shoplifting-detection platform using AWS Kinesis Video Streams, FastAPI, and a React/TypeScript monitoring interface with video overlays and real-time alert workflows.
- Combined ByteTrack/BoT-SORT tracking, MediaPipe pose and hand landmarks, trajectory deviation, dwell time, hand-object interactions, and temporal sequence scoring to detect intent beyond frame-level objects.
- Explainability: Integrated AWS Bedrock to turn multi-signal detections into human-readable intent classifications and alert explanations for reviewer validation.
- Challenges solved: Maintained identity and behavioral state across frames, reduced dependence on single noisy signals, and synchronized streaming detections with sub-second UI overlays.
- Tech: AWS Kinesis Video Streams, AWS Bedrock, CVAT, ByteTrack, BoT-SORT, MediaPipe, OpenCV, FastAPI, React, TypeScript.
- Built a microservices-based IT cost-allocation and billing platform in FastAPI/PostgreSQL supporting AS-IS/TO-BE allocation, multi-currency FX, markup snapshots, reconciliation, and auditable financial workflows.
- Developed staged Excel ingestion and high-performance XLSX reporting services with validation-before-commit for inconsistent enterprise finance datasets.
- Reliability: Used Redis/Celery for retry-safe distributed processing, automated failure recovery, idempotent job boundaries, and stable execution of long-running allocations and exports.
- Observability: Defined worker queue-depth, retry-rate, export throughput, and end-to-end p99 job-duration signals for operational dashboards and incident diagnosis.
- Controls: Added defensive validation, transactional commits, rate limiting, and audit-ready snapshots to protect high-concurrency financial writes.
- Challenges solved: Normalized irregular spreadsheets, preserved FX/markup calculations across allocation modes, and made failed background jobs safely repeatable.
- Tech: FastAPI, PostgreSQL, SQLAlchemy, Redis, Celery, Docker Compose, Pandas, OpenPyXL, XlsxWriter.
- Built an AI voice companion in Python/FastAPI integrating OpenAI (chat + speech-to-text), ElevenLabs TTS, Tavus real-time avatar video, and LiveKit sessions, with a persistent Postgres memory store injecting remembered facts and conversation history into each prompt for continuity and personalization.
- Reliability: Designed a provider-guard layer with per-user/per-provider cost caps and deterministic local fallbacks, so the app degrades cleanly (never errors out) when an AI vendor times out or a spend limit is hit.
- Safety: Engineered a safety classifier that routes emergencies to SOS steps, refuses medical diagnoses, and enforces AI-honesty guardrails in code (not just prompts) so the companion never pretends to be human - critical for a vulnerable, elderly user base.
- Shipped a cross-platform Flutter app (edge-to-edge onboarding, email-OTP auth, Google sign-in, voice UI) against a microservices backend behind a Traefik gateway with per-service PostgreSQL isolation and Alembic migrations.
- Containerized services with Docker and deployed to a reproducible cloud environment with CI running automated tests against real Postgres on every change.
- Challenges solved: Handled unreliable, metered third-party AI APIs via cost-capped provider guards with graceful fallbacks; bounded prompt context (top-N memory + recent turns) to control cost and latency; enforced non-negotiable safety routing in code rather than trusting model output.
- Tech: Python, FastAPI, OpenAI, ElevenLabs, Tavus, LiveKit, PostgreSQL, Alembic, Flutter, Dart, Docker, Traefik.
- Built data-dense React/TypeScript products for real-time sports metrics, high-velocity telemetry, CCTV event overlays, schematic editing, and drag-and-drop workflow visualization.
- Used ReactFlow, Redux Toolkit, virtualized lists, and WebSocket listeners to keep complex visual state synchronized with streaming backend events.
- Performance: Optimized render boundaries and state updates for sub-second reactivity across large datasets and frequently changing visual elements.
- Challenges solved: Prevented unnecessary re-renders, preserved graph/editor state during live updates, and unified multiple visualization patterns into reusable frontend architecture.
- Tech: React, TypeScript, ReactFlow, shadcn/ui, TailwindCSS, WebSockets, Redux Toolkit.





