WP-01 · 2026
VLMs in 2026
The Enterprise Playbook
The Enterprise Playbook
How vision-language models move from research to enterprise production. Architecture choices, evaluation, cost.
Download PDFRESOURCES · FIELD NOTES
Field notes from
shipping AI at scale.
FAQs, whitepapers, and writing from our engineering team on what actually moves an AI initiative from notebook to production.
01 How do you price engagements?
Fixed scope, fixed timeline, fixed fee per package. No T&M surprises.
02 Who owns the IP we build together?
You do. Code, fine-tuned weights, and trained models are yours.
03 Can you deploy on our infrastructure?
Yes — on-prem, your VPC, or managed SaaS. Your governance, your call.
04 How do you handle data privacy and compliance?
GDPR-compliant by default; EU AI Act conformity; India DPDP; HIPAA where the engagement requires.
05 What if we need fine-tuning vs RAG vs an off-the-shelf API?
That's exactly what the Discovery Sprint decides. We're agnostic.
06 How quickly can you get a PoC running?
4–8 weeks for most use cases, depending on data readiness.
07 Do you work with startups, or only large enterprises?
Both. Our fixed-fee model works for either.
08 What about model hallucinations and safety?
Every production system ships with evaluation harnesses, guardrails, and human-in-the-loop escalation paths.
09 Can you handle Indian languages and scripts?
Yes — we run proprietary vernacular OCR and NLU across 22 Indian languages.
10 What's your typical team composition?
ML architect + ML engineers + MLOps + product + QA. Sized to scope.
11 Do you provide ongoing support after launch?
Yes — the Managed AI Operations package covers monitoring, retraining, and roadmap.
12 How is AndOr different from other AI consultancies?
We've shipped AI to 50M+ users. Most “AI consultancies” have shipped AI to slide decks.
WHITEPAPERS
Long-form arguments
worth your time.
Each whitepaper is a self-contained POV on a hard problem — written by the engineers who run that problem in production.
WP-02 · 2026
Vernacular AI
Building NLU for India's 22 Languages
Building NLU for India's 22 Languages
Script-specific detection, Indic NLU, and the data engineering behind population-scale vernacular AI.
Download PDF
WP-03 · 2026
From PoC to Production
Why 80% of Enterprise AI Stalls
Why 80% of Enterprise AI Stalls
The operating-model failures that kill AI initiatives — and the engagement structure that gets past them.
Download PDF
WP-04 · 2026
Computer Vision Economics
Cost, Latency, Accuracy Trade-offs
Cost, Latency, Accuracy Trade-offs
The three-axis tradeoff that decides CV system design — and where consumer-scale deployment changes the math.
Download PDFFIELD NOTES
Short writing
from our engineering team.
MAR 2026
Distilling SAM for consumer phones
How we shrunk a foundation segmentation model into a 7MB student that runs in 80ms on mid-range Androids.
Read note →FEB 2026
LoRA vs full SFT: the decision rule we actually use
An opinionated frame for when LoRA is enough — and the data-volume thresholds where you should pay for full SFT.
Read note →FEB 2026
Indic OCR: conjuncts, ligatures, and what breaks vanilla CRNNs
A tour of why generic OCR underperforms on Devanagari and Bengali — and the architecture choices that fix it.
Read note →JAN 2026
Designing eval harnesses for VLMs
Why hallucination, refusal, and modality leakage need separate evals — and how we structure them in production.
Read note →DEC 2025
RAG isn't a model. It's a system.
Most RAG failures are retrieval failures, chunking failures, or eval failures — not LLM failures. A field debug guide.
Read note →DEC 2025
Generative cost engineering at consumer ARPU
The scheduler tricks, model cascades, and quantization choices that let us serve 5M designs/month profitably.
Read note →NEXT STEP