← All portfolio + Portfolio · AI Research

Portfolio · AI Research · Applied Method, Real Findings

Fine-tuning · Evaluation · Benchmarking · RAG · Agents

Applied AI research on real models & real workloads.

An applied AI research cut from Creative Mantra, the Mumbai-based AI and software agency, covering the research question, method, experiments, and findings behind real models and systems we built and evaluated. Framed the way we'd write it up for a shipping team, not a paper reviewer.

Research studies

Every AI/ML product has a fine-tune, eval or benchmark layer

Case studies in this cut

AI/ML practice · framed as applied research

Benchmarking axes

Latency · cost · accuracy — measured on real workloads

Publication style

Applied

Method → experiments → findings → what we'd study next

+ Filter by category All portfolio Web Mobile Design AIAI Research

5 featured · AI Research

What we ship in AI research

Applied research on shipped models.

An applied AI research study from Creative Mantra, the Mumbai-based AI and software agency, covering the research question, method, experiments, and findings behind a real model or system we built and evaluated.

01
Research question and problem framing.
02
Method — literature/data review, experiment design, evaluation approach.
03
Experiments, benchmarks and model/data facts.
04
Key findings and observations.
05
Conclusion, limitations and what we'd study next.

Featured studies · Method, Not Marketing

Five research cuts of the AI/ML practice.

Five applied studies — a long-context author benchmark, a multi-model persona routing eval, a detection-safe rewriting policy, an agent-in-workflow shadow-mode evaluation, and a clinician-signable computer-vision study.

See all 45+ case studies →

STUDY · 01 Research · Long-context LLM 2026

Manuscripts.ai

Manuscripts.ai · Gemini 2.5 Pro at 1M context

How does a 1M-context model behave when the entire manuscript is in the prompt?

Research question: can an author edit a 300K-word novel with an inline AI assistant that sees the whole book without hallucinating character voice? Method: 12 inline commands benchmarked on Gemini 2.5 Pro at 1M context, with a provenance ledger logging every before/after edit. Finding: the "Pad" — a brainstorm surface that refuses to output prose — is what actually keeps the manuscript clean. A top author hit 300K words with zero AI calls in the manuscript.

1M context Provenance ledger Long-form Author-first eval

Read the study →

STUDY · 02 Research · Multi-model routing 2025

Quikest

Quikest · Persona bench · Multi-model workspace

Which model is right for which persona interview — and how do you route in production?

Research question: how do you build a workspace where researchers get consistent persona answers across GPT, Gemini and Claude? Method: a persona bench, a structured-interview harness, and evals scored by consistency across sessions and models. Finding: routing per persona (not per query) produces the most stable research signal — and a UX that treats the model as a teammate, not a chatbox.

Multi-model routing Persona benchmarking Consistency evals Structured research

Read the study →

STUDY · 03 Research · Adversarial detection 2025

AI Humanize

AI Humanize · Detection-safe rewriting

Can you humanise machine-generated text without stripping the writer's voice?

Research question: what does "humanised" mean when the reader is a detector, not a person? Method: rewrites benchmarked against multiple AI detectors, with a gauge showing before/after probability and a highlights view showing which sentences changed. Finding: preserving the writer's cadence beats aggressive rewriting — the highest-scoring humanisations are the ones that leave the voice alone.

Detection benchmark Cadence-preserving Rewrite policy Writer voice

Read the study →

STUDY · 04 Research · Agent-in-workflow 2024 – 2025

T-TMS Agentic Mailer

T-TMS · Freight brokerage · Agentic mailer

How do you put an agent inside a workflow a dispatcher already trusts?

Research question: can an LLM + OCR agent read freight documents (BoLs, PODs, rate confs), identify the load and draft a reply — without making the dispatcher retrain? Method: shadow-mode evaluation on live emails, with human-in-the-loop scoring and a one-click send only after approval. Finding: the trust curve is faster when the agent shows its work — attachments identified, load matched, draft reviewed — than when it acts silently.

Agentic OCR + LLM Human-in-the-loop Shadow-mode eval

Read the study →

STUDY · 05 Research · Clinical computer vision 2025

DentalAuto

DentalAuto · Vision model tuned on intra-oral imagery

Can a vision model produce a clinician-signable finding on caries and periodontal signals?

Research question: how do you build a vision model whose output the dentist will actually sign? Method: a tuned CV model on intra-oral imagery, evaluated against expert annotation with per-signal precision/recall — packaged inside the dental PMS the clinician already uses. Finding: the model earns trust when it flags the finding on the same imagery the dentist is already looking at, in the same visit.

Computer vision Clinical eval PMS-integrated Per-signal metrics

Read the study →

Now Accepting New Research Engagements

Want to commission applied AI research?

Tell us the research question — we'll scope method, evaluation and a shipping-team write-up. Send a paragraph to hello@creative-mantra.com. Reply within 24 hours; fee letter within four working days.

Start a Study View All Work