Anonymised under NDA.

The brand still looked like a 2018 seed-stage startup. Analyst reports called the identity “generic”. Buyers shortlisted incumbents in pitch decks before even taking the demo.

1. 01Strategy refresh anchored on a single positioning sentence the CEO could defend live in front of the board.
2. 02Type system, palette and a motion language for the logo plus core UI primitives — same character across film, web, and product.
3. 03Brand guidelines + Figma library + Lottie/Rive motion kits handed to the in-house team with a maintenance playbook.

Figma · After Effects · Lottie · Rive · Frontify

Demo-to-pipeline conversion

+27%

Aided brand recall (buyer survey)

+18pp

Analyst sentiment turn

neutral → positive in two cycles

Static creative fatigued in four days. Agency turnaround sat at nine days per variant. CAC was up 38% year-on-year and the growth team was rate-limited by the studio.

1. 01In-house creative engine: modular brand-locked templates × AI-generated variants × an A/B test framework that decided what shipped.
2. 02Brand-trained image adapter for product shots; a CLIP-similarity gate so nothing off-brand reached the ad account.
3. 03Weekly creative council: human-in-the-loop review of top performers, retraining the variant generator on what actually converted.

ComfyUI + Flux LoRA · Meta Ads API · Looker · Custom React tool

Creative turnaround

9d → 8h

CAC (blended)

−23%

Weekly variant volume

14× prior

A new EV trim was launching into a three-rival fight. No Super Bowl budget, no top-of-funnel awareness time. The brief was: stand out, sell out, in fourteen days.

1. 01A single narrative thread that scaled from a 60s hero film down to 6s bumpers — no second campaign, one story.
2. 0232 dealer-localisable cuts produced from the same master sequence so regional teams shipped without re-edits.
3. 03Programmatic OOH paired with social — same beat, same line, same frame in every channel.

DaVinci Resolve · After Effects · Vimeo OTT · Vistar (OOH)

Reservation list

filled in 14 days

Organic share-of-voice

+52% during launch

Dealer-localised cuts shipped

32 in 9 days

Thought-leadership output was inconsistent across 14 markets. The brand voice was thinning out. Fintechs were outranking the firm for high-intent search queries.

1. 01Codified voice + style guide; AI-assisted drafting tool calibrated against twelve years of the firm’s own published archive.
2. 02Programmatic SEO topic graph mapped to where the firm had a defensible point of view — not where the volume was.
3. 03Editorial review gates and a LinkedIn distribution rhythm that put senior partners on every piece they shipped.

Notion · Letterdrop · Surfer SEO · LinkedIn Sales Nav

Organic article traffic

+163%

LinkedIn impressions

4.2× prior

Top-3 SERP, high-intent queries

4 articles

Compliance analysts lost 9 hours a week searching policy and regulation across six systems. New joiners took eight months to handle a case unsupervised. Audit risk was a board topic.

1. 01Hybrid retrieval (BM25 + dense embeddings + cross-encoder rerank) over 380k policy and regulation documents, with paragraph-level citation.
2. 02Citation-or-refuse contract: the model is structurally blocked from answering without three independent sources.
3. 03Eval harness of 820 expert-curated Q/A; deployment gated on faithfulness ≥ 0.92 and zero-hallucination on a 60-item red-team set.

Claude 3.5 Sonnet · pgvector · LangGraph · FastAPI · Datadog

Research time per query

−78%

Onboarding to first solo case

8 mo → 3 mo

Faithfulness on held-out eval

0.94

A legacy rule engine missed 31% of CNP fraud while blocking 8% of legitimate sales. Every false positive lost a customer; every miss landed in chargeback queue.

1. 01Streaming feature store with sub-second freshness over Kafka — same features online and offline.
2. 02Ensemble of gradient-boosted trees and a graph neural network on payment-device-merchant relationships; calibrated to a strict 12ms p99 inference budget.
3. 03Drift monitor with automated retraining loop and shadow-mode rollouts gated on PR-AUC and ECE.

Kafka · Feast · XGBoost + PyTorch Geometric · Triton · MLflow

Fraud capture rate

+29pp

False-positive rate

−64%

Annualised loss avoided

$42M

Unplanned downtime cost $7M a quarter. Vibration, current and temperature data streamed in from thousands of motors and went straight to cold storage.

1. 01Edge-trained anomaly detectors on bearing/temperature/current signals — one model per asset class, deployed via ONNX to plant gateways.
2. 02Remaining-useful-life forecasts feeding a maintenance optimiser that respects spares inventory and crew schedules.
3. 03Floor dashboard with confidence-gated alerts; every override logged for retraining.

PyTorch · ONNX · Azure IoT Edge · Databricks · Power BI

Unplanned downtime

−41%

MTBF improvement

+37%

Payback period

4.5 months

The tier-1 contact centre handled 2.4M tickets a month. 38% of agent handle time was simple intent triage. CSAT had been flat for two years.

1. 01Multi-agent orchestration: intent router → knowledge agent → action agent → human handoff, each with strict tool contracts and traced decisions.
2. 02Tool-calling against verified billing and provisioning APIs only — no free-text actions on customer accounts.
3. 03Red-team adversarial eval set covering jailbreaks, social engineering and regional policy variance; gradual rollout per market gated on resolution and complaint deltas.

LangGraph · GPT-4o · Twilio · Salesforce Service Cloud · Helicone

Self-serve resolution

+44pp

Average handle time

−3.2 min

CSAT (90-day)

+9 pts