Design for Six Sigma — embedding quality at concept stage to eliminate rework, reduce warranty risk, and achieve right-first-time outcomes every time.
Quality problems are cheap to fix in the design phase and ruinous to fix in the field, and every engineering organisation knows this. The reason it keeps happening anyway is structural: design teams are measured on release dates, manufacturing inherits variation it had no voice in specifying, and the statistical work that would have connected the two never got done because nobody owned it. Design for Six Sigma is the discipline that puts customer requirements, transfer functions, variation budgets and measurement capability into the design phase, where they cost something to establish and almost nothing to change.
This practice works with industrial manufacturers — high-voltage equipment, railway systems, energy storage, cable, process plant and heavy capital goods — in India and Germany. As a design for six sigma consultant working in India, the firm's position is that DFSS earns its place when a product family has a recurring, quantifiable defect cost or a launch that cannot afford a field escape. It does not earn its place as a training programme detached from a live product. Every deployment described below is anchored to a specific programme with a specific defect or capability target.
DMADV applies to new design; DMAIC applies to an existing process whose performance is unsatisfactory but whose fundamental design is sound. The distinction matters commercially, because DMAIC on a product whose specification is unachievable will consume months and conclude that the specification is unachievable. A short feasibility screen at the outset — is the current process capable of the tolerance at all, and does a capable process exist anywhere for this feature? — routes the work correctly.
Tollgate criteria are written as pass or fail, with the evidence named. A tollgate that can be passed by presenting a slide deck is not a tollgate.
QFD fails in practice for a predictable reason: teams build a House of Quality with 60 customer requirements against 80 technical characteristics, populate 4,800 cells with 9/3/1 relationship weights on the basis of opinion, and produce a ranking that nobody believes. The output is then filed and the design proceeds as it would have anyway.
We run QFD tightly. Customer requirements are consolidated to a manageable set and classified against the Kano model, because must-be requirements, one-dimensional requirements and delighters justify different design effort and different tolerance. Relationship weights are challenged — a 9 must be defensible by a stated mechanism, not a feeling. The correlation roof is used for what it is actually good at: exposing technical characteristics that trade against each other, which is where the real design decisions live. Where the product warrants it, the cascade runs from product requirements to part characteristics to process parameters to production controls, and the value of the cascade is that a control-plan line item can be traced back to a customer requirement. If it cannot be traced, either the requirement is missing or the control is unnecessary.
MSA comes first, always. Process data collected through an incapable gauge produces capability indices that are wrong in a direction nobody can predict, and the resulting improvement effort chases measurement noise. This is the most frequently skipped step in Indian and German shops alike, and it is the cheapest to fix.
Where a gauge fails, the corrective action is specified concretely: fixturing change, datum scheme change, operator method standardisation, or gauge replacement with the required resolution stated (a rule of ten against tolerance, relaxed only with justification).
Capability indices are widely quoted and widely misused. Three disciplines separate a useful capability study from a number on a PPAP form.
First, stability precedes capability. Cpk computed on a process that is not in statistical control describes nothing repeatable, because the estimate of within-subgroup variation is not representative of the process that produced the parts. Control charts come first; capability follows.
Second, the distinction between Cp/Cpk and Pp/Ppk is not academic. Cpk uses within-subgroup variation and describes potential; Ppk uses total variation and describes what the customer actually receives. A large gap between them is a signal about between-subgroup shift — tool wear, batch-to-batch material variation, shift changes — and that signal is more useful than either index alone.
Third, the confidence interval is rarely stated and always material. With 50 measurements, a point estimate of Cpk = 1.33 carries a 95 per cent confidence interval of roughly 1.05 to 1.61. Declaring that a 1.33 requirement has been met on 50 parts is therefore a statement with real uncertainty in it, and where a customer requires 1.67 for a safety-related characteristic, the sample size needed to demonstrate it is a planning input, not an afterthought. Non-normal data is handled by an identified distribution or a Box–Cox or Johnson transformation, or by the percentile method in the ISO 22514 framework — never by computing Cpk on skewed data and hoping.
Parameter design comes before tolerance design. Finding an operating point where the response is insensitive to noise is free; buying insensitivity through tighter tolerances is not. Our preference is a combined-array response surface design with noise factors included as model terms, and dual modelling of mean and variance, rather than the crossed inner/outer orthogonal array approach, which is expensive in runs and gives a signal-to-noise ratio that is hard to interpret when the mean and the variance move together.
Design selection is stated explicitly and justified. Full factorial 2k where k is small and all interactions matter. Fractional factorials with the resolution named — Resolution IV confounds two-factor interactions with each other and is a screening tool, not a modelling tool; Resolution V is required before two-factor interactions can be interpreted individually. Centre points are included to test for curvature before committing to a response surface. Where optimisation is the goal, central composite designs (with the axial distance chosen for rotatability or for face-centred practicality when factor levels are physically constrained) or Box–Behnken where extreme corner combinations are infeasible. Power and sample size are calculated before the runs are booked, so that a null result means something.
| Method | Assumption | When it is the right choice |
|---|---|---|
| Worst case (arithmetic) | All contributors simultaneously at their limits | Safety-critical fits, very low volumes, assemblies where a single failure is unacceptable and cost of tightening is tolerable |
| RSS (root sum of squares) | Contributors independent, approximately normal, centred, tolerance at ±3σ | Medium to high volume with demonstrated process centring; commonly with an inflation factor around 1.5 to cover non-centred and non-normal reality |
| Monte Carlo | Distributions specified per contributor; handles non-normal, skewed, non-linear and geometric relationships | Complex 3D stacks, GD&T with material condition modifiers, any stack where the assembly function is non-linear in the contributors |
Whichever method is used, the deliverable that changes behaviour is the sensitivity ranking: which three contributors own the majority of assembly variation, and which tolerances can therefore be opened up without consequence. Opening non-critical tolerances is usually the fastest cost reduction available in a mature product. Stack-ups are built on a stated datum reference frame under ASME Y14.5 or ISO GPS, with bonus tolerance from maximum material condition accounted for where the drawing permits it — a stack-up that ignores available bonus tolerance will report a problem that does not exist.
A control plan is only as good as the reaction plan attached to it. We build prototype, pre-launch and production control plans in the APQP structure, with each control traced to a failure mode in the PFMEA and each PFMEA line carrying a prevention control and a detection control that are actually different things. Risk prioritisation follows the AIAG-VDA Action Priority logic rather than a risk priority number product, because RPN thresholds create the well-known distortion of teams tuning detection ratings to fall below the action threshold.
For corrective action systems, the discipline is in the causal chain. A five-why analysis is complete only when each step is verifiable and reversible — if the stated cause is removed and reinstated, the effect should disappear and return. Where that test cannot be run, the "cause" is a hypothesis and should be labelled one. We separate containment from correction from prevention, insist on escape-point analysis (why did the control system not detect this?) alongside occurrence-point analysis, and require effectiveness verification at a defined interval after closure. An 8D closed without effectiveness data is an 8D that will reopen.
Deploying first-time-right across a product family is an organisational problem wearing a statistical costume. The technical content is design rules, DFM and DFA reviews with manufacturing present, capability data fed back from the shop floor into design tolerances, and gated release criteria. The organisational content is measurement: first-pass yield per stage, rolled throughput yield across the line, and cost of poor quality split into internal failure, external failure, appraisal and prevention. When prevention spend rises and external failure spend falls, the programme is working; when total COPQ is flat, it is not, and that should be visible in month four rather than month eighteen.
We start with an assessment against a live programme: current defect and warranty data, existing MSA and capability evidence, the FMEA and control plan as they stand, and the tolerance basis for the two or three characteristics that generate most of the cost. That produces a scoped plan with named deliverables and a target stated in the client's own units — ppm, first-pass yield, scrap value, warranty cost per unit. Work is done alongside the client's engineers rather than for them, because a DFSS capability that leaves when the consultant leaves has not been deployed. Where a client wants internal Green Belt and Black Belt capability, project work and coaching run in parallel on the same programme.
NaraNova Tech LLP was incorporated in May 2026 and is registered with DPIIT under Startup India (DIPP271026), holds Udyam registration UDYAM-GJ-24-0239773, and is GST registered. Engagements can be structured as fixed-scope projects, retained capability building, or embedded support through a product launch.
Typical outputs are a validated measurement system with the MSA study behind it, a capability baseline with confidence intervals stated, a transfer function or response surface model for the critical characteristics, a tolerance stack-up with sensitivity ranking and a specific set of tolerances recommended to tighten or open, a PFMEA and control plan with reaction plans, and a closed-loop corrective action procedure with effectiveness verification built in. All statistical work is handed over with the raw data and the analysis files, so the client can reproduce and extend it.
This site uses only cookies that are strictly necessary for it to function. We set no analytics or marketing cookies unless you accept them. See our Cookie Policy and Privacy Policy.