AI-FIRST · IN PRODUCTION SINCE 2015
From Computer Vision and Vision-Language Models to Generative AI, LLMs, and the full post-training alignment stack — AndOr turns frontier AI into deployed, measurable enterprise systems. The same proprietary models that power LightX Editor and PhotoCut, available as fixed-scope, fixed-fee services.
PROOF
Every capability we sell is already running — across consumer apps with tens of millions of users, and inside banks, hospitals, and city traffic systems.
fraud model · live banking · bfsi
Multimodal KYC, policy-grounded LLM copilot, and real-time fraud risk scoring deployed inside the bank's VPC. Sub-200ms decision latency on transaction streams. 0.91 AUC on held-out fraud cohort.
CAPABILITIES — MULTIMODAL · LLM · MLOPS
itms · live · smart cities
Real-time vehicle compliance, lane-discipline scoring, and adaptive signal timing across 40+ intersections. Edge-deployed CV pipelines running under 80ms per frame on Jetson-class hardware.
CAPABILITIES — COMPUTER VISION · EDGE · MLOPS
Consumer-scale proof: the same models, shipped to 50M+ users.
LightX Editor
50M+ Android · 5M+ iPhone
GenAI image stack at consumer cost and latency, globally.
PhotoCut
30M+ cutouts / month
CV segmentation at consumer latency, on-device and at the edge.
Photoleaf
5M designs / month
Generative pipelines for brand-consistent creative at high volume.
WHAT WE DO
End-to-end build, fine-tune, integrate, deploy, and operate AI systems across Computer Vision, Multimodal/VLM, Generative AI, LLMs, NLP, and Vernacular Intelligence.
Explore services →Maturity assessments, use-case prioritization, model and architecture selection, governance, safety, and ROI modeling — by people who've shipped AI to 50M users.
Explore consultancy →CORE EXPERTISE
Six domains where we've shipped — with the proprietary models, evaluation harnesses, and inference infrastructure to back it up.
Scene reasoning, visual Q&A, multimodal agents.
Detection, segmentation, tracking, OCR, defect inspection.
Text-to-image/video, virtual try-on, brand-consistent creative.
Indic-script OCR and NLU. 22 Indian languages, global scripts.
Domain fine-tuning, retrieval pipelines, multi-agent workflows.
Extraction, classification, summarization, enterprise search.
We don't just fine-tune. We instrument the training run, align the model, compress it for production, and keep it aligned in operation — because in 2026, that's where production value and production risk both live.
TRAINING · ALIGNMENT · OPERATIONS
Most AI partners stop at fine-tuning. We instrument the training run, align the model, compress it for production, and keep it aligned in operation. The five disciplines below are where production value — and production risk — actually live in 2026.
training observability · w&b · mlflow · on-prem
Long pretraining and continued-pretraining runs fail silently. By the time loss curves look wrong, you've burned a week of A100 hours on a broken checkpoint. We instrument the run — gradient norms, token efficiency, hardware utilization, divergence detection — and intervene before the GPU bill outruns the experiment.
CAPABILITIES — TRAINING · OBSERVABILITY · MLOPS
rlhf · dpo · orpo · constitutional ai
A fine-tuned model knows your domain. An aligned model behaves the way your business needs it to — helpful, safe, on-brand, and compliant — even on inputs it has never seen. We build the full alignment stack: preference data pipelines, reward models, and the optimization passes that turn a capable model into a deployable one.
CAPABILITIES — ALIGNMENT · COMPRESSION · SAFETY
continued pretraining · synthetic data · slm
Some workloads can't be served by an API call or a LoRA on a frontier model. Regulated industries, sub-50ms inference budgets, sovereign data constraints, and rare-domain corpora all push toward custom models. We run the entire training lifecycle — from raw corpus to deployed checkpoint — including the small specialized models that often out-perform much larger general ones in production.
CAPABILITIES — PRETRAINING · SLM · SOVEREIGN AI
hitl · preference data · continuous alignment
Alignment isn't a phase you finish. Models drift, user expectations shift, and what counted as "helpful" last quarter doesn't this one. We build the human-in-the-loop platforms that make ongoing preference collection a routine operating discipline — not a project you re-run every year.
CAPABILITIES — HITL · ALIGNMENT · MANAGED OPS
drift signals · auto-retrain · lineage
Most MLOps stops at deploy. The model goes out, dashboards go up, and the next training cycle starts from a clean slate. We close the loop: every drift signal, preference disagreement, and production failure flows back into the training data pipeline with full lineage, so the next checkpoint is informed by what production actually saw.
CAPABILITIES — MLOPS · LINEAGE · CLOSED LOOP
What's included, who each discipline is for, and how it bundles into a Production Build or Managed Operations engagement.
WHY ANDOR
50M+ downloads is non-negotiable proof — our AI runs at consumer cost and latency.
Computer Vision isn't an add-on — it's the discipline our founder began at UC Berkeley in 2002.
22 Indian languages and global scripts. Indic OCR and NLU running in production, not in research papers.
No T&M surprises. We refund the sprint if we don't deliver a written, decision-ready output.
HOW WE ENGAGE
Four steps from problem to operating production system. Predictable cost, predictable timeline, decision-ready outputs.
1–3 WEEKS
AI maturity audit, use-case scoring, build-vs-buy recommendation, indicative roadmap, and Training Readiness Assessment for teams considering custom models.
3–8 WEEKS
Working PoC against measurable acceptance criteria. Evaluation harness, cost/latency benchmark, and a written alignment plan if the deployment needs one.
8–20 WEEKS
Production-deployed AI with guardrails, monitoring, MLOps, mid-training observability, and a post-training alignment pass before go-live. Full integration into your stack.
ONGOING
Monitoring, drift detection, continuous preference collection, alignment refresh, retraining triggers, capability roadmap, and quarterly business reviews.
EVERY PHASE IS A FIXED-FEE PACKAGE · NO T&M SURPRISES
PACKAGES
Six named, fixed-fee engagements covering the full lifecycle — from initial discovery through post-training alignment and ongoing managed operations. Full comparison on services.html.
PACKAGE
Maturity audit + use-case backlog + build-vs-buy + indicative roadmap.
PACKAGE
Working PoC against acceptance criteria, evaluation harness, cost/latency benchmark.
PACKAGE · POPULAR
Production-deployed AI with guardrails, monitoring, MLOps, and stack integration.
PACKAGE
SLA-backed pod for monitoring, drift, retraining, safety, and roadmap.
PACKAGE
Real-time dashboards, divergence detection, intervention playbook, and on-prem or W&B/MLflow integration for your next pretraining or continued-pretraining run.
PACKAGE
RLHF or DPO pipeline, reward model, alignment evaluation harness, and a written alignment delta report for a production-ready, behavior-tuned model.
Fixed scope. Fixed timeline. Fixed fee. We refund the sprint if you don't get a written, decision-ready output.
IMPACT
Consumer-scale proof, measured in app stores — not in slide decks.
ANDROID DOWNLOADS
iPHONE DOWNLOADS
DESIGNS GENERATED
ENGINEERS
AI-FIRST SINCE
FOUNDER
Sharad Shankar
Founder & CEO
IIT Kanpur
2002
UC Berkeley
Auto. nav.
Punchh
Acquired
IIT Kanpur (2002) · UC Berkeley · ex-Times Internet · Punchh (acquired)
Two decades building production AI at the intersection of academia, enterprise, and startups. Today he leads a 50-engineer team building the multimodal models that power LightX, PhotoCut, and a growing book of enterprise AI engagements.
INDUSTRIES
Vertical knowledge, not vertical templates. The same proprietary models adapted to your domain's data, latency, and compliance.
CASE STUDIES
Five active engagements across government, smart cities, agritech, education, and conservation — each architected to deploy at population scale.
BANKING · BFSI
Multimodal KYC, policy-grounded LLM copilot, real-time fraud risk.
HEALTHCARE
Radiology triage, clinical-note GenAI, multilingual patient agents.
INSURANCE
Visual damage assessment, policy-grounded adjudication, fraud scoring.
MEDIA · CREATIVE OPS
Brand-LoRA pipelines for catalog, campaign, and multilingual creative.
SMART CITIES
Adaptive signal optimisation + real-time vehicle compliance and risk scoring.
EXPLORE
Architecture, KPIs, tech stack, and downloadable proposals for every engagement.
DEPLOYMENT
Three deployment models. GDPR, EU AI Act, India DPDP, and SOC2-aligned by default. HIPAA-deployable on request.
Air-gapped, your hardware, your network. We bring the stack and the operators.
Your AWS, GCP, or Azure account. Your IAM, your KMS, your audit trail.
We run the stack. You consume API endpoints with SLAs and dashboards.
11 / NEXT STEP
