10+ years of engineering, product & sales · 9+ at IBM · 5 progressive roles.
I architect enterprise AI systems at IBM — watsonx and Cloud Pak deployments for APAC's largest banks, telcos, GSIs, and governments, from reference architecture down to the install scripts. I turn the hardest systems into stunning POCs that close deals, then own the client deployments end-to-end — orchestrating partners, IBM Expert Labs, product, and support to land them in production.
On weekends: BLE wearables with ADPCM streaming at <300ms first-audio · multilingual voice bots with in-memory graph RAG (12–15× faster than Chroma) · a clinical research engine mapping symptom ↔ gene ↔ drug correlations from biomedical literature. Same muscle, different scale — whether the budget is $5M or zero, I ship what I design.
2016–2017 · Product Developer at Axelta Systems & Loop Reality, Hyderabad — stereo-vision rack management + traffic-density billboard scoring (OpenCV); ambient-monitoring hardware for refrigerated trucks (Arduino, ESP, embedded C); Android OS mods for custom hardware; VR-cycle induction-braking control + universal BLE vital-signs module integrated with Unreal Engine.
CTO Executive Team · IBM
Data / AI Client Advocacy · IBM
Hybrid Cloud Build Team · ISA · IBM
Cloud & AI Solutions Lead · Partner Ecosystem Lead
Drop a folder — it becomes a live customer demo. Plugin auto-discovery + runtime esbuild, Orchestrate-first LLM routing (timeout, caching, retryable 503s), 3-tier AES-256-GCM credentials, 29-method pluggable IStore. Hosts 12 verticalized demos (BFSI fraud, Telecom churn, Healthcare claims, Mfg predictive-maint, Govt citizen-service, etc.) used by IBM Sales / GSI Labs APAC.
The move: treated the runtime bundle as a plugin. Dropping a folder triggers esbuild → plugin manifest → auto-mount. 12 verticalized demos, one platform, zero forks.
Oxygen / beds / blood locator. Peaked at 4,000 users/day, integrated into ICMR's COVID Assist. IBM Global Research Accomplishment.
The move: crowdsourced leads went through a dedup + staleness-decay queue before surfacing, so one fake entry couldn't poison the map; outbound rate-limited to stop hospital flooding.
Infra control plane (14 ops as MCP tools) · business workflow (vision-LLM w/ graceful fallback) · 6-agent supervisor · LLM routing fabric with retryable 503s.
The move: treated cloud infra itself as a first-class tool surface — 14 control-plane APIs wrapped as MCP so agents provision, not just advise.
Data-lake and analytics architecture for India's largest automotive manufacturer — ingesting connected-car telemetry across the fleet, powering diagnostics, predictive maintenance, and fleet operations. Live in production.
The move: designed the ingest + stream-processing + analytics tiers so the OEM could run real-time fleet ops and historical deep-dives on the same lake without forking pipelines.
Map, ephemeral events, in-event chat, tribes, nudge queue. 60+ SOPs, widget-test harness per feature, feature-first clean architecture.
The move: 3-vibe taxonomy (not 40 categories) to stop discovery fragmentation; nudge queue is capacity-aware — if you've been pinged recently, the system holds back instead of spamming.
Always-on memory trigger. VAD wakes it only when you speak, IMU double-tap gesture. Backend: WhisperX (STT + diarization) + Silero VAD + FastAPI summarizer.
The move: firmware-side VAD + 8:1 ADPCM over BLE 5.0 means pendant transmits only speech, not audio. Double-tap IMU gesture = no visible button, no social stigma.
You give the room one task; the agents self-split it, execute under shared dev tools and project rules, and cross-validate each other's patches before anything ships. Every patch, run, lock, review, and handoff lands in an append-only event log — the audit trail is the artifact. Patch/review/apply puts one agent's edit in front of another for sign-off; soft TTL locks prevent collisions; floor-control keeps the human on intent and final review.
The move: distributed intelligence with cross-validation across vendor boundaries — Claude and Codex don't call each other's APIs, they coordinate through shared room events under one rulebook. One task in; peer-reviewed multi-agent work out.
A few more worth naming — ibmcloudtoolkit: a 14-tool MCP server wrapping IBM Cloud provisioning APIs (VSIs, buckets, clusters, IAM, networks) so Orchestrate agents provision infra, not just describe it; open-source shell installers for IBM Cloud Paks (CP4D / CP4I / CP4App / CP4MCM / CP4Auto / IAF / Portworx); end-to-end MLOps on Cloud Pak for Data + OpenShift — drift detection, automated retraining, and model lifecycle across tier-1 banks (first joint CP4D + CP4S enablements in region, became internal playbook); air-gapped on-prem watsonx.ai + NVIDIA GPU-cluster deployments for APAC telcos; on-demand LLM batch-processing on NVIDIA NIM + HPC for bursty inference. 87 public GitHub repos for the rest.
14+ named external conference stages — IBM Think (×5), IBM TechXchange (×2), CYPHER (India's largest AI conference, 1,100+), IBM Dev Day (×9+, AI Track Workshop Lead), Sterling OMS User Group, Gartner India, IBM Data & AI User Group Mumbai (I organized it), Infosys External Speaker Series. 90+ total talks; max live audience 3,000+; Speaker NPS 9.6.
Judge — HackOn Agentic AI (1,140 devs, 220+ teams). SME — NatWest AI for IT Hackathon. Organized IBM Call for Code across Bangalore, Pune, Hyderabad, Odisha, Kerala. Authored IBM TechXchange 2024 lab — "Governing Private LLM Deployments."
Original research & thought leadership — differential-privacy pipeline for MaaS360 device data (analytics without exposing individual device identity); neural-pattern-mapping study for puzzle learning in mice — University of Toronto collaboration, applying ML to neuroscience data.
Full write-ups, diagrams, and demos → krishnac7.com