Carrier AI SA — Data Scale & Decision Speed Reference

System / Era	Raw Input Scale	Sensor / Event Rate	Compute Platform	Decision Output	Decision Window	The Inference Problem	Hands-On Role
DARPA / USAF 1979–1983 Alpha/Sigma HEL SDI program	10,000–20,000+ sensor data points per run 6-ft diameter lasing ring · nozzles mm apart · ~2,875–5,750 nozzle clusters × 3+ sensors each · plus optical cavity, spectrometers, gas feeds, safety interlocks, cooling	Per simulation run Real-time capture during live laser run · post-run CFD analysis · Cray 1 (s/n 5) grid: est. 100K+ computational nodes · 8MB RAM · ~160 MFLOPS · state of world compute 1979	Cray 1 (s/n 5) FORTRAN One of first 5 Cray 1s built · SDI budget = no constraint	Optimal plasma flow mixing parameters for chemical laser efficiency Predict mixing patterns before committing to physical run	Run-time + post-run Hours per simulation cycle	Physical system digital twin — primitive form. Sensor inputs → virtual model of physical system → predict emergent behavior → optimize before physical test. Same architecture as Abound fault detection. 45 years earlier.	Wrote simulation code · FORTRAN · CFD modeling · sensor data integration
Invistics 2018–2021 Acquired Wolters Kluwer 2023 · NIH SBIR $2.1M · AJHP peer-reviewed	500+ hospitals 50–200 ADCs per hospital · 100–300 medication pockets per ADC · EHR cross-reference: patient records, medication orders, administration records	1–4M+ controlled substance transaction events/day Est. 2,000–8,000 transactions/hospital/day × 500+ hospitals · 8+ EHR vendors · 3+ ADC vendors · vendor-agnostic canonical data model	Vendor-agnostic integration platform Python · ML pipeline Single model across all EHR/ADC vendor combinations	10 individuals to investigate per 5,000 staff touching drugs per hospital per day 0.2% signal from 100% noise. Daily ranked output to pharmacy directors.	Daily Detection 6–8 months faster than traditional audits	Non-intrusive inference at healthcare scale. Infer individual diversion behavior from aggregate dispense log patterns. No direct observation. Per-individual classification from population-level signal. NILM pattern, different domain.	Architect + data scientist · wrote code · 96% accuracy · 3.5 yrs production · named contributor AJHP publication
J&J MedTech 2021–2023 $27B MedTech segment · 175+ countries	Multi-division SAP/JDE feeds Ethicon · DePuy Synthes · Biosense Webster · Cerenovus · J&J Vision · multiple ERP variants per division · UDI tracking across entire global device portfolio	Cycle-based Governed transformation cadence · every business requirement accepted/validated before cycle commit · Alation semantic enforcement across all feeds	Cloudera → AWS migration mid-project S3 · Redshift · Databricks · Alation Zero production defects over 2 years during live migration	EUDAMED-compliant regulated output · FDA + EU MDR dual compliance · unified semantic layer across all divisions Foundation for 2024 NVIDIA partnership + Polyphonic AI ecosystem	Cycle + continuous Built 3–5 yrs ahead of May 2026 mandatory compliance deadline	Enterprise data foundation at regulated scale. Raw multi-source chaos → DW transformation → Alation semantic governance → S3 lakehouse gold layer. One governed foundation; any future AI use case draws from the same truth. Same pattern as Abound multi-BU architecture.	Technical BA lead · lead data steward · Alation owner · acceptance gate for all cycle deliverables · wrote transformation specs
Amgen Nov 2023–Apr 2024 AWS Bedrock partnership · AAPS Open peer-reviewed · industry first	CTD Modules 3–5 regulatory corpus Chemistry, Manufacturing & Controls · Safety · Efficacy · ICH Q3/Q8/Q9/Q10/Q12 · Veeva Vault regulatory document corpus · multi-country submission requirements	Per submission cycle First GxP-validated GenAI for FDA submissions · AWS Bedrock · multi-agent · 400+ SAFe user stories · Board approval milestone · scope expanded Module 3 → Modules 4–5 after production	AWS Bedrock · Veeva Vault RIM · Accumulus Synergy · SCA templates Same Bedrock substrate as Carrier's Tell Me More	FDA-ready CMC dossier · 40% drafting time reduction · human signatory review retained Feb 2025: first digitally generated dossier to 21 countries simultaneously (Accumulus Synergy)	Submission deadline Most regulated output environment that exists	GenAI + NLP in highest-stakes regulated environment. Multi-agent GenAI layer on structured data corpus. Human-in-loop. GxP-validated. Bedrock. Same architecture as Tell Me More — GenAI layer on top of existing data system. Built this 18 months before Carrier launched it.	Architect · Product Owner · led cross-functional team · co-authored 400+ user stories · wrote architecture specs · took to Board approval
NewsRx 2024–Present PubMed phase in production · expanding to full journal catalog	Full PubMed corpus 36M+ total articles · ~4,000 new articles/day · daily full ingest · bronze/silver/gold medallion · PostgreSQL · 2,600+ production records across 6 runs · expanding to 100K+ journal sources	Daily production pipeline Daily gold processing · military-log style edge case and strong-fit outputs · dual scoring: human clinical specialists (weeks) + AI equivalent scorers (parallel, independent) · cross-compare feeds inference engine	Self-hosted Qwen3-32B · RunPod GPU · vLLM · Dagster · Qdrant (8ms p50) · MLflow · Python 3.14 Zero managed API dependency · full inference control	Daily ranked clinical outputs · edge case flags · strong fit signals · prompt improvement recommendations from scoring disagreement Human reviewers preferred AI output 88% of time · secured MVP scope beyond contracted engagement	24 hours Daily push to human reviewers · human scoring feedback loop: weeks · AI scorer feedback loop: continuous	Production LLM pipeline with dual-harness quality governance. Idempotent pipeline. Blinded independent scoring. AI scorers built to match specialist domain judgment. Feedback loop into inference. Concept drift via HHEM + BERTScore. This is not a pilot — PubMed phase is in daily production.	Wrote code · architect · all pipeline layers · scoring harness · inference configuration · production operations
Optinosis 2024–Present 10 top cancer types · clinical AI · IEC 62304 / GxP / ALCOA+	Multimodal EHR data 10 cancer type training sets · structured + unstructured clinical records · lab results · imaging metadata · treatment histories · GxP/ALCOA+ data governance from day one	Training + validation pipeline Kubeflow/Kubernetes · PyTorch + Axolotl · DPO/LoRA fine-tuning · Unsloth · RunPod H100 SXM · Modal serverless GPU · MLflow experiment tracking · advancing to production inference	vLLM · Qwen3-32B · Modal · Kubernetes · Kubeflow Pipelines Same Kubeflow abstraction as Vertex AI Pipelines — direct GCP bridge	Clinical inference output · 10-cancer scope advancing to broader oncology coverage IEC 62304 SaMD governance · ALCOA+ data integrity · GxP quality framework throughout	Advancing to production 10 cancers → full oncology scope · same build discipline as NewsRx	Fine-tuned clinical LLM under SaMD governance. Same Kubeflow pipeline architecture as Vertex AI. Same DPO/LoRA fine-tuning stack. Regulated clinical AI from ground up. Not a research project — built to IEC 62304 production standards.	Wrote code · architect · fine-tuning stack · pipeline design · governance framework
Carrier Abound Target Role TARGET AI CoE · 30-person team · Jai leads	150,000+ connected devices 32,000+ sites · 640M+ sq ft · BACnet/MQTT/webhook feeds · per-customer data isolation · AWS IoT Core ingestion	Real-time continuous telemetry IoT Core → Kinesis → Lambda → DynamoDB/Timestream → S3 → Glue → Redshift → SageMaker → Bedrock · 7B+ kWh lifetime · 40K+ dispatches avoided	AWS full stack + GCP (WeatherNext/BigQuery/Vertex AI for HEMS) Cortex pattern on both platforms	Fault prediction · energy optimization · Tell Me More conversational AI Up to 24% energy savings · $71M customer savings · 437,900 MT CO2e prevented · $28K/yr SaaS per contract	Real-time 6–8 day advance fault warning	IoT digital twin at commercial building scale. NILM-pattern per-asset ML + Bedrock GenAI layer. Same architecture Bishop has built across DARPA (physical simulation), Invistics (behavioral inference), J&J (data foundation), Amgen (Bedrock GenAI), NewsRx (production LLM). Every layer of Abound maps to a system Bishop has built and operated.	Senior AI Solutions Architect · AI CoE contributor · extend GenAI layer · scale patterns across BUs

The 45-Year Arc — One Sentence Per Era

1979: Plasma flow simulation on Cray 1 — predict physical system behavior from sensor data. Primitive digital twin.
2017: Behavioral NLP inference from SMS patterns — nascent, rule-based, Dialogflow.
2018: Non-intrusive inference at hospital scale — NILM pattern, 96% accuracy, 3.5 yrs production.
2021: Enterprise data foundation at regulated scale — J&J, Alation, AWS lakehouse, zero defects.
2023: First GxP-validated GenAI on Bedrock — FDA submissions, multi-agent, Board approval.
2024: Self-hosted frontier LLM in daily production — Qwen3-32B, vLLM, dual scoring harness.
Next: Carrier Abound — every layer of the stack, at IoT scale, in production.

NewsRx + Optinosis — Honest Production Framing

NewsRx PubMed phase: In daily production. Full ingest, daily gold, daily outputs, dual scoring harness running. This is not a pilot.
NewsRx expansion: Scope growth — adding journal sources beyond PubMed. Production architecture already proven.
Optinosis 10 cancers: Advancing to production inference. Built to IEC 62304 SaMD standards from day one — not research-grade, production-grade discipline.
The honest frame: "Both are production-grade in their current scope and expanding. I don't prototype at production scale — the governance, idempotency, and quality harness are built in from the start."
The knock it covers: POC→MVP→production is the right build sequence. Carrier's Abound followed the same path. Tell Me More was a pilot before it was a February 2026 launch.

Scale Comparison — Say This to Jai

DARPA → Carrier: "I went from 10,000+ sensors on a 6-foot lasing ring on a Cray 1 to 150,000 IoT devices on AWS. The sensor count scaled. The architecture didn't change — inputs, virtual model, predict, optimize."
Invistics → Abound: "5,000 employees touching drugs → 10 to investigate. Daily. That's the same decision support pattern as Abound's fault prediction queue — population-level signal, per-asset actionable output."
J&J → Abound multi-BU: "I built the Alation + S3 lakehouse governance pattern at J&J across divisions with different ERPs. Carrier's multi-BU scaling problem is the same architecture at HVAC scale."
Amgen → Tell Me More: "I built Bedrock GenAI on top of a structured regulated data system in 2023. Carrier launched the same pattern — Tell Me More — in February 2026. I was 18 months ahead of this problem."

    The through-line for Jai: Every system on this table is the same problem — predict behavior of a physical or behavioral system from sensor/signal data, in a regulated or high-stakes environment, at production scale, with a human-in-the-loop governance layer. DARPA was the primitive form. Carrier Abound is the mature form. I have built every layer of the stack that Abound runs on, in production, across 45 years. I didn't arrive here by accident.
  

Carrier · AI Solutions Architect · Round 1 · Data Scale & Decision Speed Reference